[WIP/TEST] distil-wisperを削除/faster-wisperの処理を修正

config.py

@@ -812,9 +812,6 @@ class Config:
             "large-v1": "large-v1",
             "large-v2": "large-v2",
             "large-v3": "large-v3",
-            "distil-small": "distil-small",
-            "distil-medium": "distil-medium",
-            "distil-large-v2": "distil-large-v2",
         }
 
         self._MAX_MIC_ENERGY_THRESHOLD = 2000
@@ -895,7 +892,7 @@ class Config:
         }
         self._USE_TRANSLATION_FEATURE = True
         self._CTRANSLATE2_WEIGHT_TYPE = "Small"
-        self._USE_WHISPER_FEATURE = True
+        self._USE_WHISPER_FEATURE = False
         self._WHISPER_WEIGHT_TYPE = "base"
         self._SEND_MESSAGE_FORMAT = "[message]"
         self._SEND_MESSAGE_FORMAT_WITH_T = "[message]([translation])"

locales/en.yml

@@ -140,9 +140,6 @@ config_window:
     large_v1: "large_v1 model (%{capacity})"
     large_v2: "large_v2 model (%{capacity})"
     large_v3: "large_v3 model (%{capacity})"
-    distil_small: "distil-small model (%{capacity})"
-    distil_medium: "distil-medium model (%{capacity})"
-    distil_large_v2: "distil-large-v2 model (%{capacity})"
 
   deepl_auth_key:
     label: DeepL Auth Key

model.py

@@ -339,12 +339,11 @@ class Model:
             source=self.mic_audio_recorder.source,
             phrase_timeout=phase_timeout,
             max_phrases=config.INPUT_MIC_MAX_PHRASES,
-            transcription_engine=config.SELECTED_TRANSCRIPTION_ENGINE,
-            whisper_weight_type=config.WHISPER_WEIGHT_TYPE,
             root=config.PATH_LOCAL,
+            whisper_weight_type=config.WHISPER_WEIGHT_TYPE,
         )
         def sendMicTranscript():
-            mic_transcriber.transcribeAudioQueue(mic_audio_queue, config.SOURCE_LANGUAGE, config.SOURCE_COUNTRY)
+            mic_transcriber.transcribeAudioQueue(mic_audio_queue, config.SOURCE_LANGUAGE, config.SOURCE_COUNTRY, config.SELECTED_TRANSCRIPTION_ENGINE)
             message = mic_transcriber.getTranscript()
             try:
                 fnc(message)
@@ -423,12 +422,11 @@ class Model:
             source=self.speaker_audio_recorder.source,
             phrase_timeout=phase_timeout,
             max_phrases=config.INPUT_SPEAKER_MAX_PHRASES,
-            transcription_engine=config.SELECTED_TRANSCRIPTION_ENGINE,
-            whisper_weight_type=config.WHISPER_WEIGHT_TYPE,
             root=config.PATH_LOCAL,
+            whisper_weight_type=config.WHISPER_WEIGHT_TYPE,
         )
         def sendSpeakerTranscript():
-            speaker_transcriber.transcribeAudioQueue(speaker_audio_queue, config.TARGET_LANGUAGE, config.TARGET_COUNTRY)
+            speaker_transcriber.transcribeAudioQueue(speaker_audio_queue, config.TARGET_LANGUAGE, config.TARGET_COUNTRY, config.SELECTED_TRANSCRIPTION_ENGINE)
             message = speaker_transcriber.getTranscript()
             try:
                 fnc(message)

models/transcription/transcription_transcriber.py

@@ -5,7 +5,7 @@ from speech_recognition import Recognizer, AudioData, AudioFile
 from datetime import timedelta
 from pyaudiowpatch import get_sample_size, paInt16
 from .transcription_languages import transcription_lang
-from .transcription_whisper import getWhisperModel
+from .transcription_whisper import getWhisperModel, checkWhisperWeight
 
 import torch
 import numpy as np
@@ -14,7 +14,7 @@ PHRASE_TIMEOUT = 3
 MAX_PHRASES = 10
 
 class AudioTranscriber:
-    def __init__(self, speaker, source, phrase_timeout, max_phrases, transcription_engine, whisper_weight_type=None, root=None):
+    def __init__(self, speaker, source, phrase_timeout, max_phrases, root=None, whisper_weight_type=None, ):
         self.speaker = speaker
         self.phrase_timeout = phrase_timeout
         self.max_phrases = max_phrases
@@ -30,34 +30,37 @@ class AudioTranscriber:
                 "new_phrase": True,
                 "process_data_func": self.processSpeakerData if speaker else self.processSpeakerData
         }
-        self.transcription_engine = transcription_engine
-        match self.transcription_engine:
-            case "Google":
-                self.audio_recognizer = Recognizer()
-            case "Whisper":
-                self.audio_recognizer = getWhisperModel(root, whisper_weight_type)
+        if whisper_weight_type is not None and root is not None and checkWhisperWeight(root, whisper_weight_type) is True:
+            self.whisper_model = getWhisperModel(root, whisper_weight_type)
+        else:
+            self.whisper_model = None
 
-    def transcribeAudioQueue(self, audio_queue, language, country):
+    def transcribeAudioQueue(self, audio_queue, language, country, transcription_engine):
         audio, time_spoken = audio_queue.get()
         self.updateLastSampleAndPhraseStatus(audio, time_spoken)
 
         text = ''
         try:
+            # Whisperが使用できない場合はGoogle Speech-to-Textを使用する
+            if transcription_engine == "Whisper":
+                if self.whisper_model is None:
+                    transcription_engine = "Google"
+
             audio_data = self.audio_sources["process_data_func"]()
-            match self.transcription_engine:
+            match transcription_engine:
                 case "Google":
-                    text = self.audio_recognizer.recognize_google(audio_data, language=transcription_lang[language][country][self.transcription_engine])
+                    text = self.audio_recognizer.recognize_google(audio_data, language=transcription_lang[language][country][transcription_engine])
                 case "Whisper":
                     audio_data = np.frombuffer(audio_data.get_raw_data(convert_rate=16000, convert_width=2), np.int16).flatten().astype(np.float32) / 32768.0
                     if isinstance(audio_data, torch.Tensor):
                         audio_data = audio_data.detach().numpy()
-                    segments, _ = self.audio_recognizer.transcribe(
+                    segments, _ = self.whisper_model.transcribe(
                         audio_data,
                         beam_size=5,
                         temperature=0.0,
                         log_prob_threshold=-0.8,
                         no_speech_threshold=0.6,
-                        language=transcription_lang[language][country][self.transcription_engine],
+                        language=transcription_lang[language][country][transcription_engine],
                         word_timestamps=False,
                         without_timestamps=True,
                         task="transcribe",

models/transcription/transcription_whisper.py

@@ -15,9 +15,6 @@ _MODELS = {
     "large-v1": "Systran/faster-whisper-large-v1",
     "large-v2": "Systran/faster-whisper-large-v2",
     "large-v3": "Systran/faster-whisper-large-v3",
-    "distil-small": "Systran/faster-distil-whisper-small.en",
-    "distil-medium": "Systran/faster-distil-whisper-medium.en",
-    "distil-large-v2": "Systran/faster-distil-whisper-large-v2"
 }
 
 _FILENAMES = [

view.py

@@ -954,9 +954,6 @@ class View():
             config.SELECTABLE_WHISPER_WEIGHT_TYPE_DICT["large-v1"]: i18n.t("config_window.whisper_weight_type.large_v1", capacity="2.87GB"),
             config.SELECTABLE_WHISPER_WEIGHT_TYPE_DICT["large-v2"]: i18n.t("config_window.whisper_weight_type.large_v2", capacity="2.87GB"),
             config.SELECTABLE_WHISPER_WEIGHT_TYPE_DICT["large-v3"]: i18n.t("config_window.whisper_weight_type.large_v3", capacity="2.87GB"),
-            config.SELECTABLE_WHISPER_WEIGHT_TYPE_DICT["distil-small"]: i18n.t("config_window.whisper_weight_type.distil_small", capacity="319MB"),
-            config.SELECTABLE_WHISPER_WEIGHT_TYPE_DICT["distil-medium"]: i18n.t("config_window.whisper_weight_type.distil_medium", capacity="755MB"),
-            config.SELECTABLE_WHISPER_WEIGHT_TYPE_DICT["distil-large-v2"]: i18n.t("config_window.whisper_weight_type.distil_large_v2", capacity="1.41GB"),
         }
 
 # Open Webpage Functions