0.2.10

Fix warmup file behavior

DEV_NOTES.md

@@ -18,8 +18,29 @@ Decoder weights: 59110771 bytes
 Encoder weights: 15268874 bytes
 
 
+# 2. Translation: Faster model for each system
 
-# 2. SortFormer Diarization: 4-to-2 Speaker Constraint Algorithm
+## Benchmark Results
+
+Testing on MacBook M3 with NLLB-200-distilled-600M model:
+
+### Standard Transformers vs CTranslate2
+
+| Test Text | Standard Inference Time | CTranslate2 Inference Time | Speedup |
+|-----------|-------------------------|---------------------------|---------|
+| UN Chief says there is no military solution in Syria | 0.9395s | 2.0472s | 0.5x |
+| The rapid advancement of AI technology is transforming various industries | 0.7171s | 1.7516s | 0.4x |
+| Climate change poses a significant threat to global ecosystems | 0.8533s | 1.8323s | 0.5x |
+| International cooperation is essential for addressing global challenges | 0.7209s | 1.3575s | 0.5x |
+| The development of renewable energy sources is crucial for a sustainable future | 0.8760s | 1.5589s | 0.6x |
+
+**Results:**
+- Total Standard time: 4.1068s
+- Total CTranslate2 time: 8.5476s
+- CTranslate2 is slower on this system --> Use Transformers, and ideally we would have an mlx implementation.
+
+
+# 3. SortFormer Diarization: 4-to-2 Speaker Constraint Algorithm
 
 Transform a diarization model that predicts up to 4 speakers into one that predicts up to 2 speakers by mapping the output predictions.
 
@@ -67,4 +88,4 @@ ELSE:
     AS_2 ← B
 
 to finish
-```
+```

README.md

@@ -18,9 +18,9 @@ Real-time speech transcription directly to your browser, with a ready-to-use bac
 
 #### Powered by Leading Research:
 
-- [SimulStreaming](https://github.com/ufal/SimulStreaming) (SOTA 2025) - Ultra-low latency transcription with AlignAtt policy
+- [SimulStreaming](https://github.com/ufalSimulStreaming) (SOTA 2025) - Ultra-low latency transcription using [AlignAtt policy](https://arxiv.org/pdf/2305.11408)
 - [NLLB](https://arxiv.org/abs/2207.04672), ([distilled](https://huggingface.co/entai2965/nllb-200-distilled-600M-ctranslate2)) (2024) - Translation to more than 100 languages.
-- [WhisperStreaming](https://github.com/ufal/whisper_streaming) (SOTA 2023) - Low latency transcription with LocalAgreement policy
+- [WhisperStreaming](https://github.com/ufal/whisper_streaming) (SOTA 2023) - Low latency transcription using [LocalAgreement policy](https://www.isca-archive.org/interspeech_2020/liu20s_interspeech.pdf)
 - [Streaming Sortformer](https://arxiv.org/abs/2507.18446) (SOTA 2025) - Advanced real-time speaker diarization
 - [Diart](https://github.com/juanmc2005/diart) (SOTA 2021) - Real-time speaker diarization
 - [Silero VAD](https://github.com/snakers4/silero-vad) (2024) - Enterprise-grade Voice Activity Detection
@@ -42,15 +42,6 @@ pip install whisperlivekit
 ```
 > You can also clone the repo and `pip install -e .` for the latest version.
 
-
->  **FFmpeg is required** and must be installed before using WhisperLiveKit
-> 
-> | OS | How to install |
-> |-----------|-------------|
->  | Ubuntu/Debian | `sudo apt install ffmpeg` |
-> | MacOS | `brew install ffmpeg` |
-> | Windows | Download .exe from https://ffmpeg.org/download.html and add to PATH |
-
 #### Quick Start
 1. **Start the transcription server:**
    ```bash
@@ -86,11 +77,11 @@ See  **Parameters & Configuration** below on how to use them.
 **Command-line Interface**: Start the transcription server with various options:
 
 ```bash
-# Use better model than default (small)
-whisperlivekit-server --model large-v3
+# Large model and translate from french to danish
+whisperlivekit-server --model large-v3 --language fr --target-language da
 
-# Advanced configuration with diarization and language
-whisperlivekit-server --host 0.0.0.0 --port 8000 --model medium --diarization --language fr
+# Diarization and server listening on */80 
+whisperlivekit-server --host 0.0.0.0 --port 80 --model medium --diarization --language fr
 ```
 
 
@@ -137,26 +128,15 @@ async def websocket_endpoint(websocket: WebSocket):
 
 ## Parameters & Configuration
 
-An important list of parameters can be changed. But what *should* you change?
-- the `--model` size. List and recommandations [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/available_models.md)
-- the `--language`.  List [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/simul_whisper/whisper/tokenizer.py). If you use `auto`, the model attempts to detect the language automatically, but it tends to bias towards English.
-- the `--backend` ? you can switch to `--backend faster-whisper` if  `simulstreaming` does not work correctly or if you prefer to avoid the dual-license requirements.
-- `--warmup-file`, if you have one
-- `--task translate`, to translate in english
-- `--host`, `--port`, `--ssl-certfile`, `--ssl-keyfile`, if you set up a server
-- `--diarization`, if you want to use it.
-- [BETA] `--target-language`, to translate using NLLB. [118 languages available](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/translation/mapping_languages.py). If you want to translate to english, you should rather use `--task translate`, since Whisper can do it directly.
-
-### Full list of parameters :
 
 | Parameter | Description | Default |
 |-----------|-------------|---------|
-| `--model` | Whisper model size. | `small` |
-| `--language` | Source language code or `auto` | `auto` |
-| `--task` | Set to `translate` to translate to english | `transcribe` |
-| `--target-language` | [BETA] Translation language target. Ex: `fr` | `None` |
-| `--backend` | Processing backend | `simulstreaming` |
-| `--min-chunk-size` | Minimum audio chunk size (seconds) | `1.0` |
+| `--model` | Whisper model size. List and recommandations [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/available_models.md) | `small` |
+| `--language` | List [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/simul_whisper/whisper/tokenizer.py). If you use `auto`, the model attempts to detect the language automatically, but it tends to bias towards English. | `auto` |
+| `--target-language` | If sets, activates translation using NLLB. Ex: `fr`. [118 languages available](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/translation/mapping_languages.py). If you want to translate to english, you should rather use `--task translate`, since Whisper can do it directly. | `None` |
+| `--task` | Set to `translate` to translate *only* to english, using Whisper translation. | `transcribe` |
+| `--diarization` | Enable speaker identification | `False` |
+| `--backend` | Processing backend. You can switch to `faster-whisper` if  `simulstreaming` does not work correctly | `simulstreaming` |
 | `--no-vac` | Disable Voice Activity Controller | `False` |
 | `--no-vad` | Disable Voice Activity Detection | `False` |
 | `--warmup-file` | Audio file path for model warmup | `jfk.wav` |
@@ -164,8 +144,19 @@ An important list of parameters can be changed. But what *should* you change?
 | `--port` | Server port | `8000` |
 | `--ssl-certfile` | Path to the SSL certificate file (for HTTPS support) | `None` |
 | `--ssl-keyfile` | Path to the SSL private key file (for HTTPS support) | `None` |
-| `--pcm-input` | raw PCM (s16le) data is expected as input and FFmpeg will be bypassed. | `False` |
+| `--pcm-input` | raw PCM (s16le) data is expected as input and FFmpeg will be bypassed. Frontend will use AudioWorklet instead of MediaRecorder | `False` |
 
+| Translation options | Description | Default |
+|-----------|-------------|---------|
+| `--nllb-backend` | `transformers` or `ctranslate2` | `ctranslate2` |
+| `--nllb-size` | `600M` or `1.3B` | `600M` |
+
+| Diarization options | Description | Default |
+|-----------|-------------|---------|
+| `--diarization-backend` |  `diart` or `sortformer` | `sortformer` |
+| `--disable-punctuation-split` |  Disable punctuation based splits. See #214 | `False` |
+| `--segmentation-model` | Hugging Face model ID for Diart segmentation model. [Available models](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `pyannote/segmentation-3.0` |
+| `--embedding-model` | Hugging Face model ID for Diart embedding model. [Available models](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `speechbrain/spkrec-ecapa-voxceleb` |
 
 | SimulStreaming backend options | Description | Default |
 |-----------|-------------|---------|
@@ -184,25 +175,16 @@ An important list of parameters can be changed. But what *should* you change?
 | `--preload-model-count` | Optional. Number of models to preload in memory to speed up loading (set up to the expected number of concurrent users) | `1` |
 
 
+
 | WhisperStreaming backend options | Description | Default |
 |-----------|-------------|---------|
 | `--confidence-validation` | Use confidence scores for faster validation | `False` |
 | `--buffer_trimming` | Buffer trimming strategy (`sentence` or `segment`) | `segment` |
 
-| Diarization options | Description | Default |
-|-----------|-------------|---------|
-| `--diarization` | Enable speaker identification | `False` |
-| `--diarization-backend` |  `diart` or `sortformer` | `sortformer` |
-| `--disable-punctuation-split` |  Disable punctuation based splits. See #214 | `False` |
-| `--segmentation-model` | Hugging Face model ID for Diart segmentation model. [Available models](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `pyannote/segmentation-3.0` |
-| `--embedding-model` | Hugging Face model ID for Diart embedding model. [Available models](https://github.com/juanmc2005/diart/tree/main?tab=readme-ov-file#pre-trained-models) | `speechbrain/spkrec-ecapa-voxceleb` |
 
 
-> For diarization using Diart, you need access to pyannote.audio models:
-> 1. [Accept user conditions](https://huggingface.co/pyannote/segmentation) for the `pyannote/segmentation` model
-> 2. [Accept user conditions](https://huggingface.co/pyannote/segmentation-3.0) for the `pyannote/segmentation-3.0` model
-> 3. [Accept user conditions](https://huggingface.co/pyannote/embedding) for the `pyannote/embedding` model
->4. Login with HuggingFace: `huggingface-cli login`
+
+> For diarization using Diart, you need to accept user conditions [here](https://huggingface.co/pyannote/segmentation) for the `pyannote/segmentation` model, [here](https://huggingface.co/pyannote/segmentation-3.0) for the `pyannote/segmentation-3.0` model and [here](https://huggingface.co/pyannote/embedding) for the `pyannote/embedding` model. **Then**, login to HuggingFace: `huggingface-cli login`
 
 ### 🚀 Deployment Guide
 

available_models.md

@@ -1,4 +1,4 @@
-# Available model sizes:
+# Available Whisper model sizes:
 
 - tiny.en (english only)
 - tiny
@@ -70,4 +70,40 @@
 2. Limited resources or need speed? → `small` or smaller
 3. Good hardware and want best quality? → `large-v3`
 4. Need fast, high-quality transcription without translation? → `large-v3-turbo`
-5. Need translation capabilities? → `large-v2` or `large-v3` (avoid turbo)
+5. Need translation capabilities? → `large-v2` or `large-v3` (avoid turbo)
+
+
+_______________________
+
+# Translation Models and Backend
+
+**Language Support**: ~200 languages
+
+## Distilled Model Sizes Available
+
+| Model | Size | Parameters | VRAM (FP16) | VRAM (INT8) | Quality |
+|-------|------|------------|-------------|-------------|---------|
+| 600M | 2.46 GB | 600M | ~1.5GB | ~800MB | Good, understandable |
+| 1.3B | 5.48 GB | 1.3B | ~3GB | ~1.5GB | Better accuracy, context |
+
+**Quality Impact**: 1.3B has ~15-25% better BLEU scores vs 600M across language pairs.
+
+## Backend Performance
+
+| Backend | Speed vs Base | Memory Usage | Quality Loss |
+|---------|---------------|--------------|--------------|
+| CTranslate2 | 6-10x faster | 40-60% less | ~5% BLEU drop |
+| Transformers | Baseline | High | None |
+| Transformers + MPS (on Apple Silicon) | 2x faster | Medium | None |
+
+**Metrics**:
+- CTranslate2: 50-100+ tokens/sec
+- Transformers: 10-30 tokens/sec
+- Apple Silicon with MPS: Up to 2x faster than CTranslate2
+
+## Quick Decision Matrix
+
+**Choose 600M**: Limited resources, close to 0 lag
+**Choose 1.3B**: Quality matters
+**Choose Transformers**: On Apple Silicon
+

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "whisperlivekit"
-version = "0.2.9"
+version = "0.2.10"
 description = "Real-time speech-to-text with speaker diarization using Whisper"
 readme = "README.md"
 authors = [

whisperlivekit/audio_processor.py

@@ -4,11 +4,11 @@ from time import time, sleep
 import math
 import logging
 import traceback
-from whisperlivekit.timed_objects import ASRToken, Silence, Line
+from whisperlivekit.timed_objects import ASRToken, Silence, Line, FrontData, State
 from whisperlivekit.core import TranscriptionEngine, online_factory, online_diarization_factory, online_translation_factory
-from whisperlivekit.ffmpeg_manager import FFmpegManager, FFmpegState
 from whisperlivekit.silero_vad_iterator import FixedVADIterator
 from whisperlivekit.results_formater import format_output
+from whisperlivekit.ffmpeg_manager import FFmpegManager, FFmpegState
 # Set up logging once
 logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
 logger = logging.getLogger(__name__)
@@ -49,9 +49,6 @@ class AudioProcessor:
         self.bytes_per_sample = 2
         self.bytes_per_sec = self.samples_per_sec * self.bytes_per_sample
         self.max_bytes_per_sec = 32000 * 5  # 5 seconds of audio at 32 kHz
-        self.last_ffmpeg_activity = time()
-        self.ffmpeg_health_check_interval = 5
-        self.ffmpeg_max_idle_time = 10
         self.is_pcm_input = self.args.pcm_input
         self.debug = False
 
@@ -68,7 +65,7 @@ class AudioProcessor:
         self.lock = asyncio.Lock()
         self.beg_loop = None #to deal with a potential little lag at the websocket initialization, this is now set in process_audio
         self.sep = " "  # Default separator
-        self.last_response_content = ""
+        self.last_response_content = FrontData()
         
         # Models and processing
         self.asr = models.asr
@@ -78,19 +75,21 @@ class AudioProcessor:
             self.vac = FixedVADIterator(models.vac_model)
         else:
             self.vac = None
-            
-        self.ffmpeg_manager = FFmpegManager(
-            sample_rate=self.sample_rate,
-            channels=self.channels
-        )
-        
-        async def handle_ffmpeg_error(error_type: str):
-            logger.error(f"FFmpeg error: {error_type}")
-            self._ffmpeg_error = error_type
-        
-        self.ffmpeg_manager.on_error_callback = handle_ffmpeg_error
+                         
+        self.ffmpeg_manager = None
+        self.ffmpeg_reader_task = None
         self._ffmpeg_error = None
-        
+
+        if not self.is_pcm_input:
+            self.ffmpeg_manager = FFmpegManager(
+                sample_rate=self.sample_rate,
+                channels=self.channels
+            )
+            async def handle_ffmpeg_error(error_type: str):
+                logger.error(f"FFmpeg error: {error_type}")
+                self._ffmpeg_error = error_type
+            self.ffmpeg_manager.on_error_callback = handle_ffmpeg_error
+             
         self.transcription_queue = asyncio.Queue() if self.args.transcription else None
         self.diarization_queue = asyncio.Queue() if self.args.diarization else None
         self.translation_queue = asyncio.Queue() if self.args.target_language else None
@@ -98,12 +97,12 @@ class AudioProcessor:
 
         self.transcription_task = None
         self.diarization_task = None
-        self.ffmpeg_reader_task = None
         self.watchdog_task = None
         self.all_tasks_for_cleanup = []
         
         if self.args.transcription:
-            self.online = online_factory(self.args, models.asr, models.tokenizer)            
+            self.online = online_factory(self.args, models.asr, models.tokenizer)        
+            self.sep = self.online.asr.sep   
         if self.args.diarization:
             self.diarization = online_diarization_factory(self.args, models.diarization_model)
         if self.args.target_language:
@@ -113,13 +112,12 @@ class AudioProcessor:
         """Convert PCM buffer in s16le format to normalized NumPy array."""
         return np.frombuffer(pcm_buffer, dtype=np.int16).astype(np.float32) / 32768.0
 
-    async def update_transcription(self, new_tokens, buffer, end_buffer, sep):
+    async def update_transcription(self, new_tokens, buffer, end_buffer):
         """Thread-safe update of transcription with new data."""
         async with self.lock:
             self.tokens.extend(new_tokens)
             self.buffer_transcription = buffer
             self.end_buffer = end_buffer
-            self.sep = sep
             
     async def update_diarization(self, end_attributed_speaker, buffer_diarization=""):
         """Thread-safe update of diarization with new data."""
@@ -152,17 +150,16 @@ class AudioProcessor:
                 latest_end = max(self.end_buffer, self.tokens[-1].end if self.tokens else 0)
                 remaining_diarization = max(0, round(latest_end - self.end_attributed_speaker, 1))
                 
-            return {
-                "tokens": self.tokens.copy(),
-                "translated_segments": self.translated_segments.copy(),
-                "buffer_transcription": self.buffer_transcription,
-                "buffer_diarization": self.buffer_diarization,
-                "end_buffer": self.end_buffer,
-                "end_attributed_speaker": self.end_attributed_speaker,
-                "sep": self.sep,
-                "remaining_time_transcription": remaining_transcription,
-                "remaining_time_diarization": remaining_diarization
-            }
+            return State(
+                tokens=self.tokens.copy(),
+                translated_segments=self.translated_segments.copy(),
+                buffer_transcription=self.buffer_transcription,
+                buffer_diarization=self.buffer_diarization,
+                end_buffer=self.end_buffer,
+                end_attributed_speaker=self.end_attributed_speaker,
+                remaining_time_transcription=remaining_transcription,
+                remaining_time_diarization=remaining_diarization
+            )
             
     async def reset(self):
         """Reset all state variables to initial values."""
@@ -174,13 +171,15 @@ class AudioProcessor:
             self.beg_loop = time()
 
     async def ffmpeg_stdout_reader(self):
-        """Read audio data from FFmpeg stdout and process it."""
+        """Read audio data from FFmpeg stdout and process it into the PCM pipeline."""
         beg = time()
-        
         while True:
             try:
-                # Check if FFmpeg is running
-                state = await self.ffmpeg_manager.get_state()
+                if self.is_stopping:
+                    logger.info("Stopping ffmpeg_stdout_reader due to stopping flag.")
+                    break
+
+                state = await self.ffmpeg_manager.get_state() if self.ffmpeg_manager else FFmpegState.STOPPED
                 if state == FFmpegState.FAILED:
                     logger.error("FFmpeg is in FAILED state, cannot read data")
                     break
@@ -188,55 +187,41 @@ class AudioProcessor:
                     logger.info("FFmpeg is stopped")
                     break
                 elif state != FFmpegState.RUNNING:
-                    logger.warning(f"FFmpeg is in {state} state, waiting...")
-                    await asyncio.sleep(0.5)
+                    await asyncio.sleep(0.1)
                     continue
-                
+
                 current_time = time()
-                elapsed_time = math.floor((current_time - beg) * 10) / 10
-                buffer_size = max(int(32000 * elapsed_time), 4096)
+                elapsed_time = max(0.0, current_time - beg)
+                buffer_size = max(int(32000 * elapsed_time), 4096)  # dynamic read
                 beg = current_time
 
                 chunk = await self.ffmpeg_manager.read_data(buffer_size)
-                        
                 if not chunk:
-                    if self.is_stopping:
-                        logger.info("FFmpeg stdout closed, stopping.")
-                        break
-                    else:
-                        # No data available, but not stopping - FFmpeg might be restarting
-                        await asyncio.sleep(0.1)
-                        continue
-                    
+                    # No data currently available
+                    await asyncio.sleep(0.05)
+                    continue
+
                 self.pcm_buffer.extend(chunk)
                 await self.handle_pcm_data()
-                    
-                    
-                    
+
+            except asyncio.CancelledError:
+                logger.info("ffmpeg_stdout_reader cancelled.")
+                break
             except Exception as e:
                 logger.warning(f"Exception in ffmpeg_stdout_reader: {e}")
-                logger.warning(f"Traceback: {traceback.format_exc()}")
-                # Try to recover by waiting a bit
-                await asyncio.sleep(1)
-                
-                # Check if we should exit
-                if self.is_stopping:
-                    break
-        
-        logger.info("FFmpeg stdout processing finished. Signaling downstream processors.")
+                logger.debug(f"Traceback: {traceback.format_exc()}")
+                await asyncio.sleep(0.2)
+
+        logger.info("FFmpeg stdout processing finished. Signaling downstream processors if needed.")
         if self.args.transcription and self.transcription_queue:
             await self.transcription_queue.put(SENTINEL)
-            logger.debug("Sentinel put into transcription_queue.")
         if self.args.diarization and self.diarization_queue:
             await self.diarization_queue.put(SENTINEL)
-            logger.debug("Sentinel put into diarization_queue.")
         if self.args.target_language and self.translation_queue:
             await self.translation_queue.put(SENTINEL)
 
-
     async def transcription_processor(self):
         """Process audio chunks for transcription."""
-        self.sep = self.online.asr.sep
         cumulative_pcm_duration_stream_time = 0.0
         
         while True:
@@ -259,12 +244,11 @@ class AudioProcessor:
                     asr_processing_logs += f" + Silence of = {item.duration:.2f}s"
                     if self.tokens:
                         asr_processing_logs += f" | last_end = {self.tokens[-1].end} |"
-                logger.info(asr_processing_logs)
-                
-                if type(item) is Silence:
+                    logger.info(asr_processing_logs)
                     cumulative_pcm_duration_stream_time += item.duration
                     self.online.insert_silence(item.duration, self.tokens[-1].end if self.tokens else 0)
                     continue
+                logger.info(asr_processing_logs)
                 
                 if isinstance(item, np.ndarray):
                     pcm_array = item
@@ -276,7 +260,7 @@ class AudioProcessor:
                 stream_time_end_of_current_pcm = cumulative_pcm_duration_stream_time
 
                 self.online.insert_audio_chunk(pcm_array, stream_time_end_of_current_pcm)
-                new_tokens, current_audio_processed_upto = self.online.process_iter()
+                new_tokens, current_audio_processed_upto = await asyncio.to_thread(self.online.process_iter)
                 
                 # Get buffer information
                 _buffer_transcript_obj = self.online.get_buffer()
@@ -300,10 +284,10 @@ class AudioProcessor:
                 new_end_buffer = max(candidate_end_times)
                 
                 await self.update_transcription(
-                    new_tokens, buffer_text, new_end_buffer, self.sep
+                    new_tokens, buffer_text, new_end_buffer
                 )
                 
-                if new_tokens and self.args.target_language and self.translation_queue:
+                if self.translation_queue:
                     for token in new_tokens:
                         await self.translation_queue.put(token)
                         
@@ -314,6 +298,14 @@ class AudioProcessor:
                 logger.warning(f"Traceback: {traceback.format_exc()}")
                 if 'pcm_array' in locals() and pcm_array is not SENTINEL : # Check if pcm_array was assigned from queue
                     self.transcription_queue.task_done()
+        
+        if self.is_stopping:
+            logger.info("Transcription processor finishing due to stopping flag.")
+            if self.diarization_queue:
+                await self.diarization_queue.put(SENTINEL)
+            if self.translation_queue:
+                await self.translation_queue.put(SENTINEL)
+
         logger.info("Transcription processor task finished.")
 
 
@@ -328,13 +320,11 @@ class AudioProcessor:
                     logger.debug("Diarization processor received sentinel. Finishing.")
                     self.diarization_queue.task_done()
                     break
-                
-                if type(item) is Silence:
+                elif type(item) is Silence:
                     cumulative_pcm_duration_stream_time += item.duration
                     diarization_obj.insert_silence(item.duration)
                     continue
-    
-                if isinstance(item, np.ndarray):
+                elif isinstance(item, np.ndarray):
                     pcm_array = item
                 else:
                     raise Exception('item should be pcm_array') 
@@ -367,14 +357,17 @@ class AudioProcessor:
         # in the future we want to have different languages for each speaker etc, so it will be more complex.
         while True:
             try:
-                token = await self.translation_queue.get() #block until at least 1 token
-                if token is SENTINEL:
+                item = await self.translation_queue.get() #block until at least 1 token
+                if item is SENTINEL:
                     logger.debug("Translation processor received sentinel. Finishing.")
                     self.translation_queue.task_done()
                     break
+                elif type(item) is Silence:
+                    online_translation.insert_silence(item.duration)
+                    continue
                 
                 # get all the available tokens for translation. The more words, the more precise
-                tokens_to_process = [token]
+                tokens_to_process = [item]
                 additional_tokens = await get_all_from_queue(self.translation_queue)
                 
                 sentinel_found = False
@@ -385,7 +378,7 @@ class AudioProcessor:
                     tokens_to_process.append(additional_token)                
                 if tokens_to_process:
                     online_translation.insert_tokens(tokens_to_process)
-                    self.translated_segments = online_translation.process()
+                    self.translated_segments = await asyncio.to_thread(online_translation.process)
                 
                 self.translation_queue.task_done()
                 for _ in additional_tokens:
@@ -398,7 +391,7 @@ class AudioProcessor:
             except Exception as e:
                 logger.warning(f"Exception in translation_processor: {e}")
                 logger.warning(f"Traceback: {traceback.format_exc()}")
-                if 'token' in locals() and token is not SENTINEL:
+                if 'token' in locals() and item is not SENTINEL:
                     self.translation_queue.task_done()
                 if 'additional_tokens' in locals():
                     for _ in additional_tokens:
@@ -407,39 +400,26 @@ class AudioProcessor:
 
     async def results_formatter(self):
         """Format processing results for output."""
-        last_sent_trans = None
-        last_sent_diar = None
         while True:
             try:
-                ffmpeg_state = await self.ffmpeg_manager.get_state()
-                if ffmpeg_state == FFmpegState.FAILED and self._ffmpeg_error:
-                    yield {
-                        "status": "error",
-                        "error": f"FFmpeg error: {self._ffmpeg_error}",
-                        "lines": [],
-                        "buffer_transcription": "",
-                        "buffer_diarization": "",
-                        "remaining_time_transcription": 0,
-                        "remaining_time_diarization": 0
-                    }
+                # If FFmpeg error occurred, notify front-end
+                if self._ffmpeg_error:
+                    yield FrontData(
+                        status="error",
+                        error=f"FFmpeg error: {self._ffmpeg_error}"
+                    )
                     self._ffmpeg_error = None
                     await asyncio.sleep(1)
                     continue
-                
+
                 # Get current state
                 state = await self.get_current_state()
-                tokens = state["tokens"]
-                buffer_transcription = state["buffer_transcription"]
-                buffer_diarization = state["buffer_diarization"]
-                end_attributed_speaker = state["end_attributed_speaker"]
-                sep = state["sep"]
                                 
                 # Add dummy tokens if needed
-                if (not tokens or tokens[-1].is_dummy) and not self.args.transcription and self.args.diarization:
+                if (not state.tokens or state.tokens[-1].is_dummy) and not self.args.transcription and self.args.diarization:
                     await self.add_dummy_token()
                     sleep(0.5)
                     state = await self.get_current_state()
-                    tokens = state["tokens"]
                 
                 # Format output
                 lines, undiarized_text, buffer_transcription, buffer_diarization = format_output(
@@ -447,53 +427,41 @@ class AudioProcessor:
                     self.silence,
                     current_time = time() - self.beg_loop if self.beg_loop else None,
                     args = self.args,
-                    debug = self.debug
+                    debug = self.debug,
+                    sep=self.sep
                 )
                 # Handle undiarized text
                 if undiarized_text:
-                    combined = sep.join(undiarized_text)
+                    combined = self.sep.join(undiarized_text)
                     if buffer_transcription:
-                        combined += sep
-                    await self.update_diarization(end_attributed_speaker, combined)
+                        combined += self.sep
+                    await self.update_diarization(state.end_attributed_speaker, combined)
                     buffer_diarization = combined
                 
                 response_status = "active_transcription"
-                if not tokens and not buffer_transcription and not buffer_diarization:
+                if not state.tokens and not buffer_transcription and not buffer_diarization:
                     response_status = "no_audio_detected"
                     lines = []
-                elif response_status == "active_transcription" and not lines:
+                elif not lines:
                     lines = [Line(
                         speaker=1,
-                        start=state.get("end_buffer", 0),
-                        end=state.get("end_buffer", 0)        
+                        start=state.end_buffer,
+                        end=state.end_buffer
                     )]
                 
-                response = {
-                    "status": response_status,
-                    "lines": [line.to_dict() for line in lines],
-                    "buffer_transcription": buffer_transcription,
-                    "buffer_diarization": buffer_diarization,
-                    "remaining_time_transcription": state["remaining_time_transcription"],
-                    "remaining_time_diarization": state["remaining_time_diarization"] if self.args.diarization else 0
-                }
-                
-                current_response_signature = f"{response_status} | " + \
-                                           ' '.join([f"{line.speaker} {line.text}" for line in lines]) + \
-                                           f" | {buffer_transcription} | {buffer_diarization}"
-                
-                trans = state["remaining_time_transcription"]
-                diar = state["remaining_time_diarization"]
-                should_push = (
-                    current_response_signature != self.last_response_content
-                    or last_sent_trans is None
-                    or round(trans, 1) != round(last_sent_trans, 1)
-                    or round(diar, 1) != round(last_sent_diar, 1)
+                response = FrontData(
+                    status=response_status,
+                    lines=lines,
+                    buffer_transcription=buffer_transcription,
+                    buffer_diarization=buffer_diarization,
+                    remaining_time_transcription=state.remaining_time_transcription,
+                    remaining_time_diarization=state.remaining_time_diarization if self.args.diarization else 0
                 )
-                if should_push and (lines or buffer_transcription or buffer_diarization or response_status == "no_audio_detected" or trans > 0 or diar > 0):
+                                
+                should_push = (response != self.last_response_content)
+                if should_push and (lines or buffer_transcription or buffer_diarization or response_status == "no_audio_detected"):
                     yield response
-                    self.last_response_content = current_response_signature
-                    last_sent_trans = trans
-                    last_sent_diar = diar
+                    self.last_response_content = response
                 
                 # Check for termination condition
                 if self.is_stopping:
@@ -507,32 +475,32 @@ class AudioProcessor:
                         logger.info("Results formatter: All upstream processors are done and in stopping state. Terminating.")
                         return
                 
-                await asyncio.sleep(0.1)  # Avoid overwhelming the client
+                await asyncio.sleep(0.05)
                 
             except Exception as e:
                 logger.warning(f"Exception in results_formatter: {e}")
                 logger.warning(f"Traceback: {traceback.format_exc()}")
-                await asyncio.sleep(0.5)  # Back off on error
+                await asyncio.sleep(0.5)
         
     async def create_tasks(self):
         """Create and start processing tasks."""
         self.all_tasks_for_cleanup = []
         processing_tasks_for_watchdog = []
 
-        success = await self.ffmpeg_manager.start()
-        if not success:
-            logger.error("Failed to start FFmpeg manager")
-            async def error_generator():
-                yield {
-                    "status": "error", 
-                    "error": "FFmpeg failed to start. Please check that FFmpeg is installed.",
-                    "lines": [],
-                    "buffer_transcription": "",
-                    "buffer_diarization": "",
-                    "remaining_time_transcription": 0,
-                    "remaining_time_diarization": 0
-                }
-            return error_generator()
+        # If using FFmpeg (non-PCM input), start it and spawn stdout reader
+        if not self.is_pcm_input:
+            success = await self.ffmpeg_manager.start()
+            if not success:
+                logger.error("Failed to start FFmpeg manager")
+                async def error_generator():
+                    yield FrontData(
+                        status="error",
+                        error="FFmpeg failed to start. Please check that FFmpeg is installed."
+                    )
+                return error_generator()
+            self.ffmpeg_reader_task = asyncio.create_task(self.ffmpeg_stdout_reader())
+            self.all_tasks_for_cleanup.append(self.ffmpeg_reader_task)
+            processing_tasks_for_watchdog.append(self.ffmpeg_reader_task)
 
         if self.args.transcription and self.online:
             self.transcription_task = asyncio.create_task(self.transcription_processor())
@@ -549,10 +517,6 @@ class AudioProcessor:
             self.all_tasks_for_cleanup.append(self.translation_task)
             processing_tasks_for_watchdog.append(self.translation_task)
         
-        self.ffmpeg_reader_task = asyncio.create_task(self.ffmpeg_stdout_reader())
-        self.all_tasks_for_cleanup.append(self.ffmpeg_reader_task)
-        processing_tasks_for_watchdog.append(self.ffmpeg_reader_task)
-
         # Monitor overall system health
         self.watchdog_task = asyncio.create_task(self.watchdog(processing_tasks_for_watchdog))
         self.all_tasks_for_cleanup.append(self.watchdog_task)
@@ -573,15 +537,6 @@ class AudioProcessor:
                             logger.error(f"{task_name} unexpectedly completed with exception: {exc}")
                         else:
                             logger.info(f"{task_name} completed normally.")
-                
-                # Check FFmpeg status through the manager
-                ffmpeg_state = await self.ffmpeg_manager.get_state()
-                if ffmpeg_state == FFmpegState.FAILED:
-                    logger.error("FFmpeg is in FAILED state, notifying results formatter")
-                    # FFmpeg manager will handle its own recovery
-                elif ffmpeg_state == FFmpegState.STOPPED and not self.is_stopping:
-                    logger.warning("FFmpeg unexpectedly stopped, attempting restart")
-                    await self.ffmpeg_manager.restart()
                     
             except asyncio.CancelledError:
                 logger.info("Watchdog task cancelled.")
@@ -601,9 +556,14 @@ class AudioProcessor:
             if created_tasks:
                 await asyncio.gather(*created_tasks, return_exceptions=True)
             logger.info("All processing tasks cancelled or finished.")
-            await self.ffmpeg_manager.stop()
-            logger.info("FFmpeg manager stopped.")
-            if self.args.diarization and hasattr(self, 'diarization') and hasattr(self.diarization, 'close'):
+
+            if not self.is_pcm_input and self.ffmpeg_manager:
+                try:
+                    await self.ffmpeg_manager.stop()
+                    logger.info("FFmpeg manager stopped.")
+                except Exception as e:
+                    logger.warning(f"Error stopping FFmpeg manager: {e}")
+            if self.args.diarization and hasattr(self, 'dianization') and hasattr(self.diarization, 'close'):
                 self.diarization.close()
             logger.info("AudioProcessor cleanup complete.")
 
@@ -617,8 +577,13 @@ class AudioProcessor:
         if not message:
             logger.info("Empty audio message received, initiating stop sequence.")
             self.is_stopping = True
-            # Signal FFmpeg manager to stop accepting data
-            await self.ffmpeg_manager.stop()
+             
+            if self.transcription_queue:
+                await self.transcription_queue.put(SENTINEL)
+
+            if not self.is_pcm_input and self.ffmpeg_manager:
+                await self.ffmpeg_manager.stop()
+
             return
 
         if self.is_stopping:
@@ -629,6 +594,9 @@ class AudioProcessor:
             self.pcm_buffer.extend(message)
             await self.handle_pcm_data()
         else:
+            if not self.ffmpeg_manager:
+                logger.error("FFmpeg manager not initialized for non-PCM input.")
+                return
             success = await self.ffmpeg_manager.write_data(message)
             if not success:
                 ffmpeg_state = await self.ffmpeg_manager.get_state()
@@ -671,6 +639,8 @@ class AudioProcessor:
                 await self.transcription_queue.put(silence_buffer)
             if self.args.diarization and self.diarization_queue:
                 await self.diarization_queue.put(silence_buffer)
+            if self.translation_queue:
+                await self.translation_queue.put(silence_buffer)
 
         if not self.silence:
             if self.args.transcription and self.transcription_queue:

whisperlivekit/basic_server.py

@@ -18,16 +18,7 @@ args = parse_args()
 transcription_engine = None
 
 @asynccontextmanager
-async def lifespan(app: FastAPI):
-    
-    #to remove after 0.2.8
-    if args.backend == "simulstreaming" and not args.disable_fast_encoder:
-        logger.warning(f"""
-{'='*50}
-WhisperLiveKit 0.2.8 has introduced a new fast encoder feature using MLX Whisper or Faster Whisper for improved speed. Use --disable-fast-encoder to disable if you encounter issues.
-{'='*50}
-    """)
-    
+async def lifespan(app: FastAPI):    
     global transcription_engine
     transcription_engine = TranscriptionEngine(
         **vars(args),
@@ -54,7 +45,7 @@ async def handle_websocket_results(websocket, results_generator):
     """Consumes results from the audio processor and sends them via WebSocket."""
     try:
         async for response in results_generator:
-            await websocket.send_json(response)
+            await websocket.send_json(response.to_dict())
         # when the results_generator finishes it means all audio has been processed
         logger.info("Results generator finished. Sending 'ready_to_stop' to client.")
         await websocket.send_json({"type": "ready_to_stop"})
@@ -72,6 +63,11 @@ async def websocket_endpoint(websocket: WebSocket):
     )
     await websocket.accept()
     logger.info("WebSocket connection opened.")
+
+    try:
+        await websocket.send_json({"type": "config", "useAudioWorklet": bool(args.pcm_input)})
+    except Exception as e:
+        logger.warning(f"Failed to send config to client: {e}")
             
     results_generator = await audio_processor.create_tasks()
     websocket_task = asyncio.create_task(handle_websocket_results(websocket, results_generator))

whisperlivekit/core.py

@@ -4,7 +4,7 @@ try:
 except ImportError:
     from .whisper_streaming_custom.whisper_online import backend_factory
     from .whisper_streaming_custom.online_asr import OnlineASRProcessor
-from whisperlivekit.warmup import warmup_asr, warmup_online
+from whisperlivekit.warmup import warmup_asr
 from argparse import Namespace
 import sys
 
@@ -43,10 +43,12 @@ class TranscriptionEngine:
             "transcription": True,
             "vad": True,
             "pcm_input": False,
+            
             # whisperstreaming params:
             "buffer_trimming": "segment",
             "confidence_validation": False,
             "buffer_trimming_sec": 15,
+            
             # simulstreaming params:
             "disable_fast_encoder": False,
             "frame_threshold": 25,
@@ -61,10 +63,15 @@ class TranscriptionEngine:
             "max_context_tokens": None,
             "model_path": './base.pt',
             "diarization_backend": "sortformer",
+            
             # diarization params:
             "disable_punctuation_split" : False,
             "segmentation_model": "pyannote/segmentation-3.0",
-            "embedding_model": "pyannote/embedding",         
+            "embedding_model": "pyannote/embedding",  
+            
+            # translation params:
+            "nllb_backend": "ctranslate2",
+            "nllb_size": "600M"
         }
 
         config_dict = {**defaults, **kwargs}
@@ -120,7 +127,7 @@ class TranscriptionEngine:
 
             else:
                 self.asr, self.tokenizer = backend_factory(self.args)
-            warmup_asr(self.asr, self.args.warmup_file) #for simulstreaming, warmup should be done in the online class not here
+                warmup_asr(self.asr, self.args.warmup_file) #for simulstreaming, warmup should be done in the online class not here
 
         if self.args.diarization:
             if self.args.diarization_backend == "diart":
@@ -142,8 +149,7 @@ class TranscriptionEngine:
                 raise Exception('Translation cannot be set with language auto')
             else:
                 from whisperlivekit.translation.translation import load_model
-                self.translation_model = load_model([self.args.lan]) #in the future we want to handle different languages for different speakers
-            
+                self.translation_model = load_model([self.args.lan], backend=self.args.nllb_backend, model_size=self.args.nllb_size) #in the future we want to handle different languages for different speakers
         TranscriptionEngine._initialized = True
 
 
@@ -155,7 +161,6 @@ def online_factory(args, asr, tokenizer, logfile=sys.stderr):
             asr,
             logfile=logfile,
         )
-        # warmup_online(online, args.warmup_file)
     else:
         online = OnlineASRProcessor(
             asr,

whisperlivekit/ffmpeg_manager.py

@@ -7,11 +7,12 @@ import contextlib
 logger = logging.getLogger(__name__)
 logging.basicConfig(level=logging.INFO)
 
-ERROR_INSTALL_INSTRUCTIONS = """
+ERROR_INSTALL_INSTRUCTIONS = f"""
+{'='*50}
 FFmpeg is not installed or not found in your system's PATH.
-Please install FFmpeg to enable audio processing.
+Alternative Solution: You can still use WhisperLiveKit without FFmpeg by adding the --pcm-input parameter. Note that when using this option, audio will not be compressed between the frontend and backend, which may result in higher bandwidth usage.
 
-Installation instructions:
+If you want to install FFmpeg:
 
 # Ubuntu/Debian:
 sudo apt update && sudo apt install ffmpeg
@@ -25,6 +26,7 @@ brew install ffmpeg
 # 3. Add the 'bin' directory (e.g., C:\\FFmpeg\\bin) to your system's PATH environment variable.
 
 After installation, please restart the application.
+{'='*50}
 """
 
 class FFmpegState(Enum):
@@ -183,6 +185,8 @@ class FFmpegManager:
     async def _drain_stderr(self):
         try:
             while True:
+                if not self.process or not self.process.stderr:
+                    break
                 line = await self.process.stderr.readline()
                 if not line:
                     break
@@ -190,4 +194,4 @@ class FFmpegManager:
         except asyncio.CancelledError:
             logger.info("FFmpeg stderr drain task cancelled.")
         except Exception as e:
-            logger.error(f"Error draining FFmpeg stderr: {e}")
+            logger.error(f"Error draining FFmpeg stderr: {e}")

whisperlivekit/parse_args.py

@@ -20,7 +20,7 @@ def parse_args():
         help="""
         The path to a speech audio wav file to warm up Whisper so that the very first chunk processing is fast.
         If not set, uses https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav.
-        If False, no warmup is performed.
+        If empty, no warmup is performed.
         """,
     )
 
@@ -177,7 +177,7 @@ def parse_args():
         "--pcm-input",
         action="store_true",
         default=False,
-        help="If set, raw PCM (s16le) data is expected as input and FFmpeg will be bypassed."
+        help="If set, raw PCM (s16le) data is expected as input and FFmpeg will be bypassed. Frontend will use AudioWorklet instead of MediaRecorder."
     )
     # SimulStreaming-specific arguments
     simulstreaming_group = parser.add_argument_group('SimulStreaming arguments (only used with --backend simulstreaming)')
@@ -287,6 +287,20 @@ def parse_args():
         help="Optional. Number of models to preload in memory to speed up loading (set up to the expected number of concurrent instances).",
     )
 
+    simulstreaming_group.add_argument(
+        "--nllb-backend",
+        type=str,
+        default="ctranslate2",
+        help="transformers or ctranslate2",
+    )
+    
+    simulstreaming_group.add_argument(
+        "--nllb-size",
+        type=str,
+        default="600M",
+        help="600M or 1.3B",
+    )
+
     args = parser.parse_args()
     
     args.transcription = not args.no_transcription

whisperlivekit/remove_silences.py

@@ -39,7 +39,7 @@ def blank_to_silence(tokens):
                         )
                 else:
                     if silence_token: #there was silence but no more
-                        if silence_token.end - silence_token.start >= MIN_SILENCE_DURATION:
+                        if silence_token.duration() >= MIN_SILENCE_DURATION:
                             cleaned_tokens.append(
                                 silence_token
                             )

whisperlivekit/results_formater.py

@@ -46,15 +46,14 @@ def append_token_to_last_line(lines, sep, token, debug_info):
         lines[-1].text += sep + token.text + debug_info
         lines[-1].end = token.end
 
-def format_output(state, silence, current_time, args, debug):
+def format_output(state, silence, current_time, args, debug, sep):
     diarization = args.diarization
     disable_punctuation_split = args.disable_punctuation_split
-    tokens = state["tokens"]
-    translated_segments = state["translated_segments"] # Here we will attribute the speakers only based on the timestamps of the segments
-    buffer_transcription = state["buffer_transcription"]
-    buffer_diarization = state["buffer_diarization"]
-    end_attributed_speaker = state["end_attributed_speaker"]
-    sep = state["sep"]
+    tokens = state.tokens
+    translated_segments = state.translated_segments # Here we will attribute the speakers only based on the timestamps of the segments
+    buffer_transcription = state.buffer_transcription
+    buffer_diarization = state.buffer_diarization
+    end_attributed_speaker = state.end_attributed_speaker
     
     previous_speaker = -1
     lines = []
@@ -124,14 +123,33 @@ def format_output(state, silence, current_time, args, debug):
             
         append_token_to_last_line(lines, sep, token, debug_info)
     if lines and translated_segments:
-        cts_idx = 0 # current_translated_segment_idx
-        for line in lines:
-            while cts_idx < len(translated_segments):
-                ts = translated_segments[cts_idx]
-                if ts.start and ts.start >= line.start and ts.end <= line.end:
-                    line.translation += ts.text + ' '
-                    cts_idx += 1
-                else:
-                    break
-    return lines, undiarized_text, buffer_transcription, '' 
-
+        unassigned_translated_segments = []
+        for ts in translated_segments:
+            assigned = False
+            for line in lines:
+                if ts and ts.overlaps_with(line):
+                    if ts.is_within(line):
+                        line.translation += ts.text + ' '
+                        assigned = True
+                        break
+                    else:
+                        ts0, ts1 = ts.approximate_cut_at(line.end)
+                        if ts0 and line.overlaps_with(ts0):
+                            line.translation += ts0.text + ' '
+                        if ts1:
+                            unassigned_translated_segments.append(ts1)
+                        assigned = True
+                        break
+            if not assigned:
+                unassigned_translated_segments.append(ts)
+        
+        if unassigned_translated_segments:
+            for line in lines:
+                remaining_segments = []
+                for ts in unassigned_translated_segments:
+                    if ts and ts.overlaps_with(line):
+                        line.translation += ts.text + ' '
+                    else:
+                        remaining_segments.append(ts)
+                unassigned_translated_segments = remaining_segments #maybe do smth in the future about that
+    return lines, undiarized_text, buffer_transcription, ''

whisperlivekit/timed_objects.py

@@ -1,5 +1,5 @@
-from dataclasses import dataclass
-from typing import Optional
+from dataclasses import dataclass, field
+from typing import Optional, Any
 from datetime import timedelta
 
 def format_time(seconds: float) -> str:
@@ -15,6 +15,21 @@ class TimedText:
     speaker: Optional[int] = -1
     probability: Optional[float] = None
     is_dummy: Optional[bool] = False
+    
+    def overlaps_with(self, other: 'TimedText') -> bool:
+        return not (self.end <= other.start or other.end <= self.start)
+    
+    def is_within(self, other: 'TimedText') -> bool:
+        return other.contains_timespan(self)
+
+    def duration(self) -> float:
+        return self.end - self.start
+
+    def contains_time(self, time: float) -> bool:
+        return self.start <= time <= self.end
+
+    def contains_timespan(self, other: 'TimedText') -> bool:
+        return self.start <= other.start and self.end >= other.end
 
 @dataclass
 class ASRToken(TimedText):
@@ -41,6 +56,34 @@ class SpeakerSegment(TimedText):
 class Translation(TimedText):
     pass
 
+    def approximate_cut_at(self, cut_time):
+        """
+        Each word in text is considered to be of duration (end-start)/len(words in text)
+        """
+        if not self.text or not self.contains_time(cut_time):
+            return self, None
+
+        words = self.text.split()
+        num_words = len(words)
+        if num_words == 0:
+            return self, None
+
+        duration_per_word = self.duration() / num_words
+        
+        cut_word_index = int((cut_time - self.start) / duration_per_word)
+        
+        if cut_word_index >= num_words:
+            cut_word_index = num_words -1
+        
+        text0 = " ".join(words[:cut_word_index])
+        text1 = " ".join(words[cut_word_index:])
+
+        segment0 = Translation(start=self.start, end=cut_time, text=text0)
+        segment1 = Translation(start=cut_time, end=self.end, text=text1)
+
+        return segment0, segment1
+        
+
 @dataclass
 class Silence():
     duration: float
@@ -57,4 +100,38 @@ class Line(TimedText):
             'translation': self.translation,
             'start': format_time(self.start),
             'end': format_time(self.end),
-        }
+        }
+        
+@dataclass  
+class FrontData():
+    status: str = ''
+    error: str = ''
+    lines: list[Line] = field(default_factory=list)
+    buffer_transcription: str = ''
+    buffer_diarization: str = ''
+    remaining_time_transcription: float = 0.
+    remaining_time_diarization: float = 0.
+    
+    def to_dict(self):
+        _dict = {
+            'status': self.status,
+            'lines': [line.to_dict() for line in self.lines],
+            'buffer_transcription': self.buffer_transcription,
+            'buffer_diarization': self.buffer_diarization,
+            'remaining_time_transcription': self.remaining_time_transcription,
+            'remaining_time_diarization': self.remaining_time_diarization,
+        }
+        if self.error:
+            _dict['error'] = self.error
+        return _dict
+    
+@dataclass  
+class State():
+    tokens: list
+    translated_segments: list
+    buffer_transcription: str
+    buffer_diarization: str
+    end_buffer: float
+    end_attributed_speaker: float
+    remaining_time_transcription: float
+    remaining_time_diarization: float

whisperlivekit/translation/translation.py

@@ -1,3 +1,5 @@
+import logging
+import time
 import ctranslate2
 import torch
 import transformers
@@ -6,38 +8,42 @@ import huggingface_hub
 from whisperlivekit.translation.mapping_languages import get_nllb_code
 from whisperlivekit.timed_objects import Translation
 
+logger = logging.getLogger(__name__)
 
 #In diarization case, we may want to translate just one speaker, or at least start the sentences there
 
 PUNCTUATION_MARKS = {'.', '!', '?', '。', '！', '？'}
 
+MIN_SILENCE_DURATION_DEL_BUFFER = 3 #After a silence of x seconds, we consider the model should not use the buffer, even if the previous
+# sentence is not finished.
 
 @dataclass
 class TranslationModel():
     translator: ctranslate2.Translator
     tokenizer: dict
+    device: str
+    backend_type: str = 'ctranslate2'
 
-def load_model(src_langs):
-    MODEL = 'nllb-200-distilled-600M-ctranslate2'
-    MODEL_GUY = 'entai2965'
-    huggingface_hub.snapshot_download(MODEL_GUY + '/' + MODEL,local_dir=MODEL)
+def load_model(src_langs, backend='ctranslate2', model_size='600M'):
     device = "cuda" if torch.cuda.is_available() else "cpu"
-    translator = ctranslate2.Translator(MODEL,device=device)
+    MODEL = f'nllb-200-distilled-{model_size}-ctranslate2'
+    if backend=='ctranslate2':
+        MODEL_GUY = 'entai2965'
+        huggingface_hub.snapshot_download(MODEL_GUY + '/' + MODEL,local_dir=MODEL)
+        translator = ctranslate2.Translator(MODEL,device=device)
+    elif backend=='transformers':
+        translator = transformers.AutoModelForSeq2SeqLM.from_pretrained(f"facebook/nllb-200-distilled-{model_size}")
     tokenizer = dict()
     for src_lang in src_langs:
         tokenizer[src_lang] = transformers.AutoTokenizer.from_pretrained(MODEL, src_lang=src_lang, clean_up_tokenization_spaces=True)
+
     return TranslationModel(
         translator=translator,
-        tokenizer=tokenizer
+        tokenizer=tokenizer,
+        backend_type=backend,
+        device = device
     )
 
-def translate(input, translation_model, tgt_lang):
-    source = translation_model.tokenizer.convert_ids_to_tokens(translation_model.tokenizer.encode(input))
-    target_prefix = [tgt_lang]
-    results = translation_model.translator.translate_batch([source], target_prefix=[target_prefix])
-    target = results[0].hypotheses[0][1:]
-    return translation_model.tokenizer.decode(translation_model.tokenizer.convert_tokens_to_ids(target))
-
 class OnlineTranslation:
     def __init__(self, translation_model: TranslationModel, input_languages: list, output_languages: list):
         self.buffer = []
@@ -68,12 +74,19 @@ class OnlineTranslation:
             output_lang = self.output_languages[0]
         nllb_output_lang = get_nllb_code(output_lang)
             
-        source = self.translation_model.tokenizer[input_lang].convert_ids_to_tokens(self.translation_model.tokenizer[input_lang].encode(input))   
-        results = self.translation_model.translator.translate_batch([source], target_prefix=[[nllb_output_lang]]) #we can use return_attention=True to try to optimize the stuff.
-        target = results[0].hypotheses[0][1:]
-        results = self.translation_model.tokenizer[input_lang].decode(self.translation_model.tokenizer[input_lang].convert_tokens_to_ids(target))
-        return results
-
+        tokenizer = self.translation_model.tokenizer[input_lang]
+        tokenizer_output = tokenizer(input, return_tensors="pt").to(self.translation_model.device)
+        
+        if self.translation_model.backend_type == 'ctranslate2':
+            source = tokenizer.convert_ids_to_tokens(tokenizer_output['input_ids'][0])    
+            results = self.translation_model.translator.translate_batch([source], target_prefix=[[nllb_output_lang]])
+            target = results[0].hypotheses[0][1:]
+            result = tokenizer.decode(tokenizer.convert_tokens_to_ids(target))
+        else:
+            translated_tokens = self.translation_model.translator.generate(**tokenizer_output, forced_bos_token_id=tokenizer.convert_tokens_to_ids(nllb_output_lang))
+            result = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
+        return result
+    
     def translate_tokens(self, tokens):
         if tokens:
             text = ' '.join([token.text for token in tokens])
@@ -88,7 +101,6 @@ class OnlineTranslation:
             return translation
         return None
             
-        
 
     def insert_tokens(self, tokens):
         self.buffer.extend(tokens)
@@ -109,7 +121,11 @@ class OnlineTranslation:
         self.translation_remaining = self.translate_tokens(self.buffer)
         self.len_processed_buffer = len(self.buffer)
         return self.validated + [self.translation_remaining]
-                
+
+    def insert_silence(self, silence_duration: float):
+        if silence_duration >= MIN_SILENCE_DURATION_DEL_BUFFER:
+            self.buffer = []
+            self.validated += [self.translation_remaining]
 
 if __name__ == '__main__':
     output_lang = 'fr'
@@ -122,16 +138,13 @@ if __name__ == '__main__':
     test = test_string.split(' ')
     step = len(test) // 3
     
-    shared_model = load_model([input_lang])
+    shared_model = load_model([input_lang], backend='ctranslate2')
     online_translation = OnlineTranslation(shared_model, input_languages=[input_lang], output_languages=[output_lang])
-        
+    
+    beg_inference = time.time()    
     for id in range(5):
         val = test[id*step : (id+1)*step]
         val_str = ' '.join(val)
         result = online_translation.translate(val_str)
         print(result)
-    
-    
-    
-    
-    # print(result)
+    print('inference time:', time.time() - beg_inference)

whisperlivekit/warmup.py

@@ -6,57 +6,46 @@ logger = logging.getLogger(__name__)
 def load_file(warmup_file=None, timeout=5):
     import os
     import tempfile
+    import urllib.request
     import librosa
-        
+
+    if warmup_file == "":
+        logger.info(f"Skipping warmup.")
+        return None
+
+    # Download JFK sample if not already present
     if warmup_file is None:
-        # Download JFK sample if not already present
         jfk_url = "https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav"
         temp_dir = tempfile.gettempdir()
         warmup_file = os.path.join(temp_dir, "whisper_warmup_jfk.wav")
-        
-        if not os.path.exists(warmup_file):
-            logger.debug(f"Downloading warmup file from {jfk_url}")
-            print(f"Downloading warmup file from {jfk_url}")
-            import time
-            import urllib.request
-            import urllib.error
-            import socket
-            
-            original_timeout = socket.getdefaulttimeout()
-            socket.setdefaulttimeout(timeout)
-            
-            start_time = time.time()
+        if not os.path.exists(warmup_file) or os.path.getsize(warmup_file) == 0:
             try:
-                urllib.request.urlretrieve(jfk_url, warmup_file)
-                logger.debug(f"Download successful in {time.time() - start_time:.2f}s")
-            except (urllib.error.URLError, socket.timeout) as e:
-                logger.warning(f"Download failed: {e}. Proceeding without warmup.")
+                logger.debug(f"Downloading warmup file from {jfk_url}")
+                with urllib.request.urlopen(jfk_url, timeout=timeout) as r, open(warmup_file, "wb") as f:
+                    f.write(r.read())
+            except Exception as e:
+                logger.warning(f"Warmup file download failed: {e}.")
                 return None
-            finally:
-                socket.setdefaulttimeout(original_timeout)
-    elif not warmup_file:
-        return None 
-    
-    if not warmup_file or not os.path.exists(warmup_file) or os.path.getsize(warmup_file) == 0:
-        logger.warning(f"Warmup file {warmup_file} invalid or missing.")
+
+    # Validate file and load
+    if not os.path.exists(warmup_file) or os.path.getsize(warmup_file) == 0:
+        logger.warning(f"Warmup file {warmup_file} is invalid or missing.")
         return None
-    
+
     try:
-        audio, sr = librosa.load(warmup_file, sr=16000)
+        audio, _ = librosa.load(warmup_file, sr=16000)
+        return audio
     except Exception as e:
-        logger.warning(f"Failed to load audio file: {e}")
+        logger.warning(f"Failed to load warmup file: {e}")
         return None
-    return audio
 
 def warmup_asr(asr, warmup_file=None, timeout=5):
     """
     Warmup the ASR model by transcribing a short audio file.
     """
-    audio = load_file(warmup_file=None, timeout=5)
+    audio = load_file(warmup_file=warmup_file, timeout=timeout)
+    if audio is None:
+        logger.warning("Warmup file unavailable. Skipping ASR warmup.")
+        return
     asr.transcribe(audio)
-    logger.info("ASR model is warmed up")
-    
-def warmup_online(online, warmup_file=None, timeout=5):
-    audio = load_file(warmup_file=None, timeout=5)
-    online.warmup(audio)
-    logger.warning("ASR is warmed up")
+    logger.info("ASR model is warmed up.")

whisperlivekit/web/live_transcription.css

@@ -438,7 +438,6 @@ label {
   font-size: 13px;
   border-radius: 30px;
   padding: 2px 10px;
-  display: none;
 }
 
 .loading {

whisperlivekit/web/live_transcription.js

@@ -12,6 +12,8 @@ let timerInterval = null;
 let audioContext = null;
 let analyser = null;
 let microphone = null;
+let workletNode = null;
+let recorderWorker = null;
 let waveCanvas = document.getElementById("waveCanvas");
 let waveCtx = waveCanvas.getContext("2d");
 let animationFrame = null;
@@ -20,6 +22,9 @@ let lastReceivedData = null;
 let lastSignature = null;
 let availableMicrophones = [];
 let selectedMicrophoneId = null;
+let serverUseAudioWorklet = null;
+let configReadyResolve;
+const configReady = new Promise((r) => (configReadyResolve = r));
 
 waveCanvas.width = 60 * (window.devicePixelRatio || 1);
 waveCanvas.height = 30 * (window.devicePixelRatio || 1);
@@ -226,6 +231,14 @@ function setupWebSocket() {
 
     websocket.onmessage = (event) => {
       const data = JSON.parse(event.data);
+      if (data.type === "config") {
+        serverUseAudioWorklet = !!data.useAudioWorklet;
+        statusText.textContent = serverUseAudioWorklet
+          ? "Connected. Using AudioWorklet (PCM)."
+          : "Connected. Using MediaRecorder (WebM).";
+        if (configReadyResolve) configReadyResolve();
+        return;
+      }
 
       if (data.type === "ready_to_stop") {
         console.log("Ready to stop received, finalizing display and closing WebSocket.");
@@ -332,13 +345,6 @@ function renderLinesWithBuffer(
       }
 
       let currentLineText = item.text || "";
-      
-      if (item.translation) {
-        currentLineText += `<div class="label_translation">
-          <img src="/web/src/translate.svg" alt="Translation" width="12" height="12" />
-          <span>${item.translation}</span>
-        </div>`;
-      }
 
       if (idx === lines.length - 1) {
         if (!isFinalizing && item.speaker !== -2) {
@@ -372,6 +378,13 @@ function renderLinesWithBuffer(
           }
         }
       }
+      
+      if (item.translation) {
+        currentLineText += `<div class="label_translation">
+          <img src="/web/src/translate.svg" alt="Translation" width="12" height="12" />
+          <span>${item.translation}</span>
+        </div>`;
+      }
 
       return currentLineText.trim().length > 0 || speakerLabel.length > 0
         ? `<p>${speakerLabel}<br/><div class='textcontent'>${currentLineText}</div></p>`
@@ -457,13 +470,54 @@ async function startRecording() {
     microphone = audioContext.createMediaStreamSource(stream);
     microphone.connect(analyser);
 
-    recorder = new MediaRecorder(stream, { mimeType: "audio/webm" });
-    recorder.ondataavailable = (e) => {
-      if (websocket && websocket.readyState === WebSocket.OPEN) {
-        websocket.send(e.data);
+    if (serverUseAudioWorklet) {
+      if (!audioContext.audioWorklet) {
+        throw new Error("AudioWorklet is not supported in this browser");
       }
-    };
-    recorder.start(chunkDuration);
+      await audioContext.audioWorklet.addModule("/web/pcm_worklet.js");
+      workletNode = new AudioWorkletNode(audioContext, "pcm-forwarder", { numberOfInputs: 1, numberOfOutputs: 0, channelCount: 1 });
+      microphone.connect(workletNode);
+
+      recorderWorker = new Worker("/web/recorder_worker.js");
+      recorderWorker.postMessage({
+        command: "init",
+        config: {
+          sampleRate: audioContext.sampleRate,
+        },
+      });
+
+      recorderWorker.onmessage = (e) => {
+        if (websocket && websocket.readyState === WebSocket.OPEN) {
+          websocket.send(e.data.buffer);
+        }
+      };
+
+      workletNode.port.onmessage = (e) => {
+        const data = e.data;
+        const ab = data instanceof ArrayBuffer ? data : data.buffer;
+        recorderWorker.postMessage(
+          {
+            command: "record",
+            buffer: ab,
+          },
+          [ab]
+        );
+      };
+    } else {
+      try {
+        recorder = new MediaRecorder(stream, { mimeType: "audio/webm" });
+      } catch (e) {
+        recorder = new MediaRecorder(stream);
+      }
+      recorder.ondataavailable = (e) => {
+        if (websocket && websocket.readyState === WebSocket.OPEN) {
+          if (e.data && e.data.size > 0) {
+            websocket.send(e.data);
+          }
+        }
+      };
+      recorder.start(chunkDuration);
+    }
 
     startTime = Date.now();
     timerInterval = setInterval(updateTimer, 1000);
@@ -502,10 +556,28 @@ async function stopRecording() {
   }
 
   if (recorder) {
-    recorder.stop();
+    try {
+      recorder.stop();
+    } catch (e) {
+    }
     recorder = null;
   }
 
+  if (recorderWorker) {
+    recorderWorker.terminate();
+    recorderWorker = null;
+  }
+  
+  if (workletNode) {
+    try {
+      workletNode.port.onmessage = null;
+    } catch (e) {}
+    try {
+      workletNode.disconnect();
+    } catch (e) {}
+    workletNode = null;
+  }
+
   if (microphone) {
     microphone.disconnect();
     microphone = null;
@@ -549,9 +621,11 @@ async function toggleRecording() {
     console.log("Connecting to WebSocket");
     try {
       if (websocket && websocket.readyState === WebSocket.OPEN) {
+        await configReady;
         await startRecording();
       } else {
         await setupWebSocket();
+        await configReady;
         await startRecording();
       }
     } catch (err) {

whisperlivekit/web/pcm_worklet.js

@@ -0,0 +1,16 @@
+class PCMForwarder extends AudioWorkletProcessor {
+  process(inputs) {
+    const input = inputs[0];
+    if (input && input[0] && input[0].length) {
+      // Forward mono channel (0). If multi-channel, downmixing can be added here.
+      const channelData = input[0];
+      const copy = new Float32Array(channelData.length);
+      copy.set(channelData);
+      this.port.postMessage(copy, [copy.buffer]);
+    }
+    // Keep processor alive
+    return true;
+  }
+}
+
+registerProcessor('pcm-forwarder', PCMForwarder);

whisperlivekit/web/recorder_worker.js

@@ -0,0 +1,58 @@
+let sampleRate = 48000;
+let targetSampleRate = 16000;
+
+self.onmessage = function (e) {
+  switch (e.data.command) {
+    case 'init':
+      init(e.data.config);
+      break;
+    case 'record':
+      record(e.data.buffer);
+      break;
+  }
+};
+
+function init(config) {
+  sampleRate = config.sampleRate;
+  targetSampleRate = config.targetSampleRate || 16000;
+}
+
+function record(inputBuffer) {
+  const buffer = new Float32Array(inputBuffer);
+  const resampledBuffer = resample(buffer, sampleRate, targetSampleRate);
+  const pcmBuffer = toPCM(resampledBuffer);
+  self.postMessage({ buffer: pcmBuffer }, [pcmBuffer]);
+}
+
+function resample(buffer, from, to) {
+    if (from === to) {
+        return buffer;
+    }
+    const ratio = from / to;
+    const newLength = Math.round(buffer.length / ratio);
+    const result = new Float32Array(newLength);
+    let offsetResult = 0;
+    let offsetBuffer = 0;
+    while (offsetResult < result.length) {
+        const nextOffsetBuffer = Math.round((offsetResult + 1) * ratio);
+        let accum = 0, count = 0;
+        for (let i = offsetBuffer; i < nextOffsetBuffer && i < buffer.length; i++) {
+            accum += buffer[i];
+            count++;
+        }
+        result[offsetResult] = accum / count;
+        offsetResult++;
+        offsetBuffer = nextOffsetBuffer;
+    }
+    return result;
+}
+
+function toPCM(input) {
+  const buffer = new ArrayBuffer(input.length * 2);
+  const view = new DataView(buffer);
+  for (let i = 0; i < input.length; i++) {
+    const s = Math.max(-1, Math.min(1, input[i]));
+    view.setInt16(i * 2, s < 0 ? s * 0x8000 : s * 0x7FFF, true);
+  }
+  return buffer;
+}