better words in README.md

README.md

@@ -33,7 +33,7 @@ Please, cite us. [Bibtex citation](http://www.afnlp.org/conferences/ijcnlp2023/p
 
 1) ``pip install librosa`` -- audio processing library
 
-2) Whisper backend.
+2) **Whisper backend**.
 
 Two alternative backends are integrated. The most recommended one is [faster-whisper](https://github.com/guillaumekln/faster-whisper) with GPU support. Follow their instructions for NVIDIA libraries -- we succeeded with CUDNN 8.5.0 and CUDA 11.7. Install with `pip install faster-whisper`.
 
@@ -41,9 +41,16 @@ Alternative, less restrictive, but slower backend is [whisper-timestamped](https
 
 The backend is loaded only when chosen. The unused one does not have to be installed.
 
+Or: **Seamless Streaming** -- alternative to Whisper, wrapped to enable the same operation modes and input/output format. 
+
+`pip install fairseq2 pydub sentencepiece git+https://github.com/facebookresearch/seamless_communication.git`
+
+Installation suggested [here](https://github.com/facebookresearch/seamless_communication/blob/main/Seamless_Tutorial.ipynb), for special torch version cases refer to [fairseq2](https://github.com/facebookresearch/fairseq2#variants). 
+
+
 3) Optional, not recommended: sentence segmenter (aka sentence tokenizer) 
 
-Two buffer trimming options are integrated and evaluated. They have impact on
+Two buffer trimming options are integrated and evaluated for Whisper backends. They have impact on
 the quality and latency. The default "segment" option performs better according
 to our tests and does not require any sentence segmentation installed. 
 
@@ -68,8 +75,9 @@ In case of installation issues of opus-fast-mosestokenizer, especially on Window
 ### Real-time simulation from audio file
 
 ```
-usage: whisper_online.py [-h] [--min-chunk-size MIN_CHUNK_SIZE] [--model {tiny.en,tiny,base.en,base,small.en,small,medium.en,medium,large-v1,large-v2,large-v3,large}] [--model_cache_dir MODEL_CACHE_DIR] [--model_dir MODEL_DIR] [--lan LAN] [--task {transcribe,translate}]
-                         [--backend {faster-whisper,whisper_timestamped}] [--vad] [--buffer_trimming {sentence,segment}] [--buffer_trimming_sec BUFFER_TRIMMING_SEC] [--start_at START_AT] [--offline] [--comp_unaware]
+usage: whisper_online.py [-h] [--min-chunk-size MIN_CHUNK_SIZE] [--model {tiny.en,tiny,base.en,base,small.en,small,medium.en,medium,large-v1,large-v2,large-v3,large}] [--model_cache_dir MODEL_CACHE_DIR]
+                         [--model_dir MODEL_DIR] [--lan LAN] [--task {transcribe,translate}] [--backend {faster-whisper,whisper_timestamped,seamless}] [--vad] [--buffer_trimming {sentence,segment}]
+                         [--buffer_trimming_sec BUFFER_TRIMMING_SEC] [--start_at START_AT] [--offline] [--comp_unaware]
                          audio_path
 
 positional arguments:
@@ -78,24 +86,26 @@ positional arguments:
 options:
   -h, --help            show this help message and exit
   --min-chunk-size MIN_CHUNK_SIZE
-                        Minimum audio chunk size in seconds. It waits up to this time to do processing. If the processing takes shorter time, it waits, otherwise it processes the whole segment that was received by this time.
+                        Minimum audio chunk size in seconds. It waits up to this time to do processing. If the processing takes shorter time, it waits, otherwise it processes the whole segment that was received
+                        by this time. Applicable both to Whisper and seamless.
   --model {tiny.en,tiny,base.en,base,small.en,small,medium.en,medium,large-v1,large-v2,large-v3,large}
-                        Name size of the Whisper model to use (default: large-v2). The model is automatically downloaded from the model hub if not present in model cache dir.
+                        Name size of the Whisper model to use (default: large-v2). The model is automatically downloaded from the model hub if not present in model cache dir. Not applicable to seamless.
   --model_cache_dir MODEL_CACHE_DIR
-                        Overriding the default model cache dir where models downloaded from the hub are saved
+                        Overriding the default model cache dir where models downloaded from the hub are saved. Not applicable to seamless.
   --model_dir MODEL_DIR
-                        Dir where Whisper model.bin and other files are saved. This option overrides --model and --model_cache_dir parameter.
+                        Dir where Whisper model.bin and other files are saved. This option overrides --model and --model_cache_dir parameter. Not applicable to seamless.
   --lan LAN, --language LAN
-                        Language code for transcription, e.g. en,de,cs.
+                        Language code for transcription, e.g. en,de,cs. Seamless backend has its own 3-letter language codes, e.g. eng, deu, ces.
   --task {transcribe,translate}
                         Transcribe or translate.
-  --backend {faster-whisper,whisper_timestamped}
-                        Load only this backend for Whisper processing.
-  --vad                 Use VAD = voice activity detection, with the default parameters.
+  --backend {faster-whisper,whisper_timestamped,seamless}
+                        Load only this backend for Whisper processing, or Seamless Streaming.
+  --vad                 Use VAD = voice activity detection, with the default parameters. Not applicable to seamless.
   --buffer_trimming {sentence,segment}
-                        Buffer trimming strategy -- trim completed sentences marked with punctuation mark and detected by sentence segmenter, or the completed segments returned by Whisper. Sentence segmenter must be installed for "sentence" option.
+                        Buffer trimming strategy -- trim completed sentences marked with punctuation mark and detected by sentence segmenter, or the completed segments returned by Whisper. Sentence segmenter
+                        must be installed for "sentence" option. Not applicable to seamless.
   --buffer_trimming_sec BUFFER_TRIMMING_SEC
-                        Buffer trimming length threshold in seconds. If buffer length is longer, trimming sentence/segment is triggered.
+                        Buffer trimming length threshold in seconds. If buffer length is longer, trimming sentence/segment is triggered. Not applicable to seamless.
   --start_at START_AT   Start processing audio at this time.
   --offline             Offline mode.
   --comp_unaware        Computationally unaware simulation.

seamless_integration.py

@@ -0,0 +1,172 @@
+#!/usr/bin/env python3
+import sys
+import numpy as np
+
+# code extracted from https://github.com/facebookresearch/seamless_communication/blob/main/Seamless_Tutorial.ipynb :
+
+from simuleval.data.segments import SpeechSegment, EmptySegment
+from simuleval.utils.arguments import cli_argument_list
+from simuleval import options
+
+from typing import Union, List
+from simuleval.data.segments import Segment, TextSegment
+from simuleval.agents.pipeline import TreeAgentPipeline
+from simuleval.agents.states import AgentStates
+
+SAMPLE_RATE = 16000
+
+def reset_states(system, states):
+    if isinstance(system, TreeAgentPipeline):
+        states_iter = states.values()
+    else:
+        states_iter = states
+    for state in states_iter:
+        state.reset()
+
+def get_states_root(system, states) -> AgentStates:
+    if isinstance(system, TreeAgentPipeline):
+        # self.states is a dict
+        return states[system.source_module]
+    else:
+        # self.states is a list
+        return system.states[0]
+
+def build_streaming_system(model_configs, agent_class):
+    parser = options.general_parser()
+    parser.add_argument("-f", "--f", help="a dummy argument to fool ipython", default="1")
+
+    agent_class.add_args(parser)
+    args, _ = parser.parse_known_args(cli_argument_list(model_configs))
+    system = agent_class.from_args(args)
+    return system
+
+class OutputSegments:
+    def __init__(self, segments: Union[List[Segment], Segment]):
+        if isinstance(segments, Segment):
+            segments = [segments]
+        self.segments: List[Segment] = [s for s in segments]
+
+    @property
+    def is_empty(self):
+        return all(segment.is_empty for segment in self.segments)
+
+    @property
+    def finished(self):
+        return all(segment.finished for segment in self.segments)
+
+
+######################
+# fixing DetokenizerAgent -- it strips output segment.content last space, but sometimes a word is split into more segments. Simple joining with spaces would be wrong.
+from seamless_communication.streaming.agents.detokenizer import DetokenizerAgent
+from seamless_communication.streaming.agents.offline_w2v_bert_encoder import (
+    OfflineWav2VecBertEncoderAgent,
+)
+from seamless_communication.streaming.agents.online_feature_extractor import (
+    OnlineFeatureExtractorAgent,
+)
+from seamless_communication.streaming.agents.online_text_decoder import (
+    MMASpeechToTextDecoderAgent,
+)
+from seamless_communication.streaming.agents.silero_vad import SileroVADAgent
+from seamless_communication.streaming.agents.unity_pipeline import UnitYAgentPipeline
+class FixDetokenizerAgent(DetokenizerAgent):
+    def decode(self, x: str) -> str:
+        return x.replace(" ", "").replace("\u2581", " ")  # .strip() is removed
+
+class FixSeamlessStreamingS2TVADAgent(UnitYAgentPipeline):
+    pipeline = [
+        SileroVADAgent,
+        OnlineFeatureExtractorAgent,
+        OfflineWav2VecBertEncoderAgent,
+        MMASpeechToTextDecoderAgent,
+        FixDetokenizerAgent,
+    ]
+##################################
+
+# the next pieces of are copypasted from the tutorial and put to the corresponding methods
+
+#class SeamlessProcessor(OnlineASRProcessorBase):  # TODO: there should be a common base class. But the code would not be simple anymore.
+class SeamlessProcessor:
+    '''
+    Wrapping SeamlessStreaming for the same operation modes as
+    Whisper-Streaming's OnlineASRProcessor. 
+
+    '''
+    def __init__(self, tgt_lan, task, logfile=sys.stderr):
+        '''
+        tgt_lan:    must be 3-letter language code that Seamless-Streaming supports for text output mode.
+        task:   see below
+        logfile
+        '''
+        if task in ("transcribe","asr"):
+            task_arg = "asr"
+        elif task in ("translate","s2tt"):
+            task_arg = "s2tt"
+        else:
+            raise ValueError("task argument must be 'transcribe' or 'translate', or 'asr' or 's2tt'")
+
+        self.logfile = logfile
+
+        agent_class = FixSeamlessStreamingS2TVADAgent
+
+        model_configs = dict(
+            source_segment_size=320,
+            device="cuda:0",
+            dtype="fp16",
+            min_starting_wait_w2vbert=192,
+            decision_threshold=0.5,
+            min_unit_chunk_size=50,
+            no_early_stop=True,
+            max_len_a=0,
+            max_len_b=100,
+            task=task_arg,
+            tgt_lang=tgt_lan,
+            block_ngrams=True,
+            detokenize_only=True,
+        )
+        self.tgt_lan = tgt_lan
+
+        self.system = build_streaming_system(model_configs, agent_class)
+
+        self.system_states = self.system.build_states()
+
+        self.init()
+
+    def init(self):
+        reset_states(self.system, self.system_states)
+        self.audio_buffer = np.array([],dtype=np.float32)
+        self.beg, self.end = 0, 0
+
+    def insert_audio_chunk(self, audio):
+        self.audio_buffer = np.append(self.audio_buffer, audio)
+
+    def process_segment(self, input_segment):
+        output_segments = OutputSegments(self.system.pushpop(input_segment, self.system_states))
+        out = []
+        for segment in output_segments.segments:
+            if not segment.is_empty:
+                out.append(segment.content)
+        if output_segments.finished:
+            print("End of VAD segment",file=self.logfile)
+            reset_states(self.system, self.system_states)
+        if out:
+            b = self.beg
+            self.beg = self.end
+            o = "".join(out)
+            return (b, self.end, "".join(out))
+        return (None, None, "")
+
+
+    def process_iter(self, finished=False):
+        input_segment = SpeechSegment(
+                content=self.audio_buffer,
+                sample_rate=SAMPLE_RATE,
+                finished=finished,
+        )
+        self.audio_buffer = np.array([],dtype=np.float32)
+        input_segment.tgt_lang = self.tgt_lan
+        self.end += (len(input_segment.content)/SAMPLE_RATE)
+        return self.process_segment(input_segment)
+
+    def finish(self):
+        return self.process_iter(finished=True)

whisper_online.py

@@ -208,7 +208,18 @@ class HypothesisBuffer:
     def complete(self):
         return self.buffer
 
-class OnlineASRProcessor:
+class OnlineASRProcessorBase:
+    '''Showing minimum common public interface for various specialized subclasses.'''
+    def init(self):
+        raise NotImplemented()
+    def insert_audio_chunk(self, audio):
+        raise NotImplemented()
+    def process_iter(self):
+        raise NotImplemented()
+    def finish(self):
+        raise NotImplemented()
+
+class OnlineASRProcessor(OnlineASRProcessorBase):
 
     SAMPLING_RATE = 16000
 
@@ -410,6 +421,7 @@ class OnlineASRProcessor:
             e = offset + sents[-1][1]
         return (b,e,t)
 
+
 WHISPER_LANG_CODES = "af,am,ar,as,az,ba,be,bg,bn,bo,br,bs,ca,cs,cy,da,de,el,en,es,et,eu,fa,fi,fo,fr,gl,gu,ha,haw,he,hi,hr,ht,hu,hy,id,is,it,ja,jw,ka,kk,km,kn,ko,la,lb,ln,lo,lt,lv,mg,mi,mk,ml,mn,mr,ms,mt,my,ne,nl,nn,no,oc,pa,pl,ps,pt,ro,ru,sa,sd,si,sk,sl,sn,so,sq,sr,su,sv,sw,ta,te,tg,th,tk,tl,tr,tt,uk,ur,uz,vi,yi,yo,zh".split(",")
 
 def create_tokenizer(lan):
@@ -447,16 +459,16 @@ def add_shared_args(parser):
     """shared args for simulation (this entry point) and server
     parser: argparse.ArgumentParser object
     """
-    parser.add_argument('--min-chunk-size', type=float, default=1.0, help='Minimum audio chunk size in seconds. It waits up to this time to do processing. If the processing takes shorter time, it waits, otherwise it processes the whole segment that was received by this time.')
-    parser.add_argument('--model', type=str, default='large-v2', choices="tiny.en,tiny,base.en,base,small.en,small,medium.en,medium,large-v1,large-v2,large-v3,large".split(","),help="Name size of the Whisper model to use (default: large-v2). The model is automatically downloaded from the model hub if not present in model cache dir.")
-    parser.add_argument('--model_cache_dir', type=str, default=None, help="Overriding the default model cache dir where models downloaded from the hub are saved")
-    parser.add_argument('--model_dir', type=str, default=None, help="Dir where Whisper model.bin and other files are saved. This option overrides --model and --model_cache_dir parameter.")
-    parser.add_argument('--lan', '--language', type=str, default='en', help="Language code for transcription, e.g. en,de,cs.")
+    parser.add_argument('--min-chunk-size', type=float, default=1.0, help='Minimum audio chunk size in seconds. It waits up to this time to do processing. If the processing takes shorter time, it waits, otherwise it processes the whole segment that was received by this time. Applicable both to Whisper and seamless.')
+    parser.add_argument('--model', type=str, default='large-v2', choices="tiny.en,tiny,base.en,base,small.en,small,medium.en,medium,large-v1,large-v2,large-v3,large".split(","),help="Name size of the Whisper model to use (default: large-v2). The model is automatically downloaded from the model hub if not present in model cache dir. Not applicable to seamless.")
+    parser.add_argument('--model_cache_dir', type=str, default=None, help="Overriding the default model cache dir where models downloaded from the hub are saved. Not applicable to seamless.")
+    parser.add_argument('--model_dir', type=str, default=None, help="Dir where Whisper model.bin and other files are saved. This option overrides --model and --model_cache_dir parameter. Not applicable to seamless.")
+    parser.add_argument('--lan', '--language', type=str, default='en', help="Language code for transcription, e.g. en,de,cs. Seamless backend has its own 3-letter language codes, e.g. eng, deu, ces.")
     parser.add_argument('--task', type=str, default='transcribe', choices=["transcribe","translate"],help="Transcribe or translate.")
-    parser.add_argument('--backend', type=str, default="faster-whisper", choices=["faster-whisper", "whisper_timestamped"],help='Load only this backend for Whisper processing.')
-    parser.add_argument('--vad', action="store_true", default=False, help='Use VAD = voice activity detection, with the default parameters.')
-    parser.add_argument('--buffer_trimming', type=str, default="segment", choices=["sentence", "segment"],help='Buffer trimming strategy -- trim completed sentences marked with punctuation mark and detected by sentence segmenter, or the completed segments returned by Whisper. Sentence segmenter must be installed for "sentence" option.')
-    parser.add_argument('--buffer_trimming_sec', type=float, default=15, help='Buffer trimming length threshold in seconds. If buffer length is longer, trimming sentence/segment is triggered.')
+    parser.add_argument('--backend', type=str, default="faster-whisper", choices=["faster-whisper", "whisper_timestamped", "seamless"],help='Load only this backend for Whisper processing, or Seamless Streaming.')
+    parser.add_argument('--vad', action="store_true", default=False, help='Use VAD = voice activity detection, with the default parameters. Not applicable to seamless.')
+    parser.add_argument('--buffer_trimming', type=str, default="segment", choices=["sentence", "segment"],help='Buffer trimming strategy -- trim completed sentences marked with punctuation mark and detected by sentence segmenter, or the completed segments returned by Whisper. Sentence segmenter must be installed for "sentence" option. Not applicable to seamless.')
+    parser.add_argument('--buffer_trimming_sec', type=float, default=15, help='Buffer trimming length threshold in seconds. If buffer length is longer, trimming sentence/segment is triggered. Not applicable to seamless.')
 
 ## main:
 
@@ -488,45 +500,50 @@ if __name__ == "__main__":
     size = args.model
     language = args.lan
 
+    min_chunk = args.min_chunk_size
+    
+    if args.backend != "seamless":
+        # loading Whisper model
         t = time.time()
         print(f"Loading Whisper {size} model for {language}...",file=logfile,end=" ",flush=True)
 
         if args.backend == "faster-whisper":
             asr_cls = FasterWhisperASR
-    else:
+        elif args.backend == "whisper_timestamped":
             asr_cls = WhisperTimestampedASR
 
         asr = asr_cls(modelsize=size, lan=language, cache_dir=args.model_cache_dir, model_dir=args.model_dir)
 
-    if args.task == "translate":
-        asr.set_translate_task()
-        tgt_language = "en"  # Whisper translates into English
-    else:
-        tgt_language = language  # Whisper transcribes in this language
-
-
         e = time.time()
         print(f"done. It took {round(e-t,2)} seconds.",file=logfile)
 
         if args.vad:
             print("setting VAD filter",file=logfile)
             asr.use_vad()
+        if args.task == "translate":
+            asr.set_translate_task()
+            tgt_language = "en"  # Whisper translates into English
+        else:
+            tgt_language = language  # Whisper transcribes in this language
 
-    
-    min_chunk = args.min_chunk_size
         if args.buffer_trimming == "sentence":
             tokenizer = create_tokenizer(tgt_language)
         else:
             tokenizer = None
+
         online = OnlineASRProcessor(asr,tokenizer,logfile=logfile,buffer_trimming=(args.buffer_trimming, args.buffer_trimming_sec))
-
-
         # load the audio into the LRU cache before we start the timer
         a = load_audio_chunk(audio_path,0,1)
 
         # warm up the ASR, because the very first transcribe takes much more time than the other
         asr.transcribe(a)
 
+    else:
+        print(f"Loading Seamless Streaming backend model",file=logfile,flush=True)
+
+        from seamless_integration import SeamlessProcessor
+        online = SeamlessProcessor(language, args.task, logfile=logfile)
+
     beg = args.start_at
     start = time.time()-beg
 

whisper_online_server.py

@@ -20,8 +20,11 @@ args = parser.parse_args()
 
 SAMPLING_RATE = 16000
 
-size = args.model
 language = args.lan
+min_chunk = args.min_chunk_size
+
+if args.backend != "seamless":  # loading Whisper backend
+    size = args.model
 
     t = time.time()
     print(f"Loading Whisper {size} model for {language}...",file=sys.stderr,end=" ",flush=True)
@@ -50,17 +53,6 @@ if args.vad:
         print("setting VAD filter",file=sys.stderr)
         asr.use_vad()
 
-
-min_chunk = args.min_chunk_size
-
-if args.buffer_trimming == "sentence":
-    tokenizer = create_tokenizer(tgt_language)
-else:
-    tokenizer = None
-online = OnlineASRProcessor(asr,tokenizer,buffer_trimming=(args.buffer_trimming, args.buffer_trimming_sec))
-
-
-
     demo_audio_path = "cs-maji-2.16k.wav"
     if os.path.exists(demo_audio_path):
         # load the audio into the LRU cache before we start the timer
@@ -72,8 +64,16 @@ if os.path.exists(demo_audio_path):
     else:
         print("Whisper is not warmed up",file=sys.stderr)
 
+    if args.buffer_trimming == "sentence":
+        tokenizer = create_tokenizer(tgt_language)
+    else:
+        tokenizer = None
+    online = OnlineASRProcessor(asr,tokenizer,buffer_trimming=(args.buffer_trimming, args.buffer_trimming_sec))
+else:  # seamless backend:
+    print(f"Loading Seamless Streaming backend model",file=sys.stderr,flush=True)
 
-
+    from seamless_integration import SeamlessProcessor
+    online = SeamlessProcessor(language, args.task, logfile=sys.stderr)
 
 ######### Server objects