LLM/WhisperLiveKit
1
0
Fork 0
mirror of https://github.com/QuentinFuxa/WhisperLiveKit.git synced 2026-03-07 22:33:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

sep for mlx is also ""

Browse Source
This commit is contained in:
Silas Kieser
2025-01-21 10:39:38 +01:00
parent ce56264241
commit 7fad0a3ee2
2 changed files with 13 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/whisper_streaming/backends.py
View File

@@ -164,7 +164,7 @@ class MLXWhisper(ASRBase):
Significantly faster than faster-whisper (without CUDA) on Apple M1.
"""
sep = " "
sep = "" # In my experience in french it should also be no space.
def load_model(self, modelsize=None, cache_dir=None, model_dir=None):
"""

15
src/whisper_streaming/online_asr.py
View File

@@ -166,7 +166,13 @@ class OnlineASRProcessor:
if (
len(self.audio_buffer) / self.SAMPLING_RATE > self.buffer_trimming_sec
): # longer than this
logger.debug("chunking sentence")
self.chunk_completed_sentence()
else:
logger.debug("not enough audio to trim as a sentence")
if self.buffer_trimming_way == "segment":
s = self.buffer_trimming_sec # trim the completed segments longer than s,
@@ -194,7 +200,10 @@ class OnlineASRProcessor:
def chunk_completed_sentence(self):
if self.commited == []:
return
logger.debug("COMPLETED SENTENCE: ", [s[2] for s in self.commited])
import pdb; pdb.set_trace()
raw_text = self.asr.sep.join([s[2] for s in self.commited])
logger.debug(f"COMPLETED SENTENCE: {raw_text}")
sents = self.words_to_sentences(self.commited)
for s in sents:
logger.debug(f"\t\tSENT: {s}")
@@ -243,7 +252,7 @@ class OnlineASRProcessor:
"""
cwords = [w for w in words]
t = " ".join(o[2] for o in cwords)
t = self.asr.sep.join(o[2] for o in cwords)
s = self.tokenize(t)
out = []
while s:
@@ -269,7 +278,7 @@ class OnlineASRProcessor:
"""
o = self.transcript_buffer.complete()
f = self.to_flush(o)
logger.debug(f"last, noncommited: {f}")
logger.debug(f"last, noncommited: {f[0]*1000:.0f}-{f[1]*1000:.0f}: {f[2]}")
self.buffer_time_offset += len(self.audio_buffer) / 16000
return f