SilasKieser commited on
Commit
b18d80f
·
1 Parent(s): 937fc50

sep for mlx is also ""

Browse files
src/whisper_streaming/backends.py CHANGED
@@ -164,7 +164,7 @@ class MLXWhisper(ASRBase):
164
  Significantly faster than faster-whisper (without CUDA) on Apple M1.
165
  """
166
 
167
- sep = " "
168
 
169
  def load_model(self, modelsize=None, cache_dir=None, model_dir=None):
170
  """
 
164
  Significantly faster than faster-whisper (without CUDA) on Apple M1.
165
  """
166
 
167
+ sep = "" # In my experience in french it should also be no space.
168
 
169
  def load_model(self, modelsize=None, cache_dir=None, model_dir=None):
170
  """
src/whisper_streaming/online_asr.py CHANGED
@@ -175,7 +175,13 @@ class OnlineASRProcessor:
175
  if (
176
  len(self.audio_buffer) / self.SAMPLING_RATE > self.buffer_trimming_sec
177
  ): # longer than this
 
 
178
  self.chunk_completed_sentence()
 
 
 
 
179
 
180
  if self.buffer_trimming_way == "segment":
181
  s = self.buffer_trimming_sec # trim the completed segments longer than s,
@@ -286,7 +292,7 @@ class OnlineASRProcessor:
286
  """
287
  o = self.transcript_buffer.complete()
288
  f = self.to_flush(o)
289
- logger.debug(f"last, noncommited: {f}")
290
  self.buffer_time_offset += len(self.audio_buffer) / 16000
291
  return f
292
 
 
175
  if (
176
  len(self.audio_buffer) / self.SAMPLING_RATE > self.buffer_trimming_sec
177
  ): # longer than this
178
+
179
+ logger.debug("chunking sentence")
180
  self.chunk_completed_sentence()
181
+
182
+
183
+ else:
184
+ logger.debug("not enough audio to trim as a sentence")
185
 
186
  if self.buffer_trimming_way == "segment":
187
  s = self.buffer_trimming_sec # trim the completed segments longer than s,
 
292
  """
293
  o = self.transcript_buffer.complete()
294
  f = self.to_flush(o)
295
+ logger.debug(f"last, noncommited: {f[0]*1000:.0f}-{f[1]*1000:.0f}: {f[2]}")
296
  self.buffer_time_offset += len(self.audio_buffer) / 16000
297
  return f
298