ailia-speech 1.3.0.5__py3-none-any.whl → 1.3.1.0__py3-none-any.whl

ailia_speech/__init__.py

@@ -67,6 +67,7 @@ AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_SMALL = (2)
 AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_MEDIUM = (3)
 AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE = (4)
 AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3 = (5)
+AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3_TURBO = (6)
 
 AILIA_SPEECH_TASK_TRANSCRIBE = (0)
 AILIA_SPEECH_TASK_TRANSLATE = (1)
@@ -425,6 +426,12 @@ class Whisper(AiliaSpeechModel):
             decoder_path = "decoder_large_v3_fix_kv_cache.onnx"
             encoder_pb_path = "encoder_large_v3_weights.pb"
             decoder_pb_path = "decoder_large_v3_fix_kv_cache_weights.pb"
+        elif model_type == AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3_TURBO:
+            encoder_path = "encoder_turbo.onnx"
+            decoder_path = "decoder_turbo_fix_kv_cache.onnx"
+            encoder_pb_path = "encoder_turbo_weights.pb"
+            decoder_pb_path = None
+            model_type = AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3
         self._download_model(model_path, encoder_path, decoder_path, encoder_pb_path, decoder_pb_path)
         self._open_model(model_path + encoder_path, model_path + decoder_path, model_type)
         self._open_vad(model_path + "silero_vad.onnx", AILIA_SPEECH_VAD_TYPE_SILERO)
@@ -494,6 +501,43 @@ class Whisper(AiliaSpeechModel):
 
         self._check(dll.ailiaSpeechResetTranscribeState(self._instance))
 
+    def transcribe_step(self, audio_waveform, sampling_rate, complete, lang = None):
+        if len(audio_waveform.shape) == 1:
+            channels = 1
+        elif len(audio_waveform.shape) == 2:
+            channels = audio_waveform.shape[0]
+            audio_waveform = numpy.transpose(audio_waveform, (1, 0)).flatten()
+        else:
+            raise AiliaSpeechError(f"audio_waveform must be 1 channel or 2 channel", -1)
+
+        audio_waveform = numpy.ascontiguousarray(audio_waveform.astype(numpy.float32))
+
+        if lang is not None:
+            self._check(dll.ailiaSpeechSetLanguage(self._instance, self._string_buffer(lang)))
+
+        self._check(dll.ailiaSpeechPushInputData(self._instance, audio_waveform, channels, audio_waveform.shape[0] // channels, sampling_rate))
+        if complete:
+            self._check(dll.ailiaSpeechFinalizeInputData(self._instance))
+
+        while True:
+            buffered = ctypes.c_uint(0)
+            self._check(dll.ailiaSpeechBuffered(self._instance, ctypes.byref(buffered)))
+            if buffered.value == 0:
+                break
+
+            self._check(dll.ailiaSpeechTranscribe(self._instance))
+
+            count = ctypes.c_uint(0)
+            self._check(dll.ailiaSpeechGetTextCount(self._instance, ctypes.byref(count)))
+            results = []
+            for i in range(count.value):
+                text = AILIASpeechText()
+                self._check(dll.ailiaSpeechGetText(self._instance, ctypes.byref(text), AILIA_SPEECH_TEXT_VERSION, i))
+                yield {"text" : text.text.decode(), "time_stamp_begin" : text.time_stamp_begin, "time_stamp_end" : text.time_stamp_end, "person_id" : text.person_id, "language" : text.language.decode(), "confidence" : text.confidence}
+
+        if complete:
+            self._check(dll.ailiaSpeechResetTranscribeState(self._instance))
+
     def __del__(self):
         if self._instance:
             dll.ailiaSpeechDestroy(cast(self._instance, c_void_p))

{ailia_speech-1.3.0.5.data → ailia_speech-1.3.1.0.data}/scripts/__init__.py

@@ -67,6 +67,7 @@ AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_SMALL = (2)
 AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_MEDIUM = (3)
 AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE = (4)
 AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3 = (5)
+AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3_TURBO = (6)
 
 AILIA_SPEECH_TASK_TRANSCRIBE = (0)
 AILIA_SPEECH_TASK_TRANSLATE = (1)
@@ -425,6 +426,12 @@ class Whisper(AiliaSpeechModel):
             decoder_path = "decoder_large_v3_fix_kv_cache.onnx"
             encoder_pb_path = "encoder_large_v3_weights.pb"
             decoder_pb_path = "decoder_large_v3_fix_kv_cache_weights.pb"
+        elif model_type == AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3_TURBO:
+            encoder_path = "encoder_turbo.onnx"
+            decoder_path = "decoder_turbo_fix_kv_cache.onnx"
+            encoder_pb_path = "encoder_turbo_weights.pb"
+            decoder_pb_path = None
+            model_type = AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3
         self._download_model(model_path, encoder_path, decoder_path, encoder_pb_path, decoder_pb_path)
         self._open_model(model_path + encoder_path, model_path + decoder_path, model_type)
         self._open_vad(model_path + "silero_vad.onnx", AILIA_SPEECH_VAD_TYPE_SILERO)
@@ -494,6 +501,43 @@ class Whisper(AiliaSpeechModel):
 
         self._check(dll.ailiaSpeechResetTranscribeState(self._instance))
 
+    def transcribe_step(self, audio_waveform, sampling_rate, complete, lang = None):
+        if len(audio_waveform.shape) == 1:
+            channels = 1
+        elif len(audio_waveform.shape) == 2:
+            channels = audio_waveform.shape[0]
+            audio_waveform = numpy.transpose(audio_waveform, (1, 0)).flatten()
+        else:
+            raise AiliaSpeechError(f"audio_waveform must be 1 channel or 2 channel", -1)
+
+        audio_waveform = numpy.ascontiguousarray(audio_waveform.astype(numpy.float32))
+
+        if lang is not None:
+            self._check(dll.ailiaSpeechSetLanguage(self._instance, self._string_buffer(lang)))
+
+        self._check(dll.ailiaSpeechPushInputData(self._instance, audio_waveform, channels, audio_waveform.shape[0] // channels, sampling_rate))
+        if complete:
+            self._check(dll.ailiaSpeechFinalizeInputData(self._instance))
+
+        while True:
+            buffered = ctypes.c_uint(0)
+            self._check(dll.ailiaSpeechBuffered(self._instance, ctypes.byref(buffered)))
+            if buffered.value == 0:
+                break
+
+            self._check(dll.ailiaSpeechTranscribe(self._instance))
+
+            count = ctypes.c_uint(0)
+            self._check(dll.ailiaSpeechGetTextCount(self._instance, ctypes.byref(count)))
+            results = []
+            for i in range(count.value):
+                text = AILIASpeechText()
+                self._check(dll.ailiaSpeechGetText(self._instance, ctypes.byref(text), AILIA_SPEECH_TEXT_VERSION, i))
+                yield {"text" : text.text.decode(), "time_stamp_begin" : text.time_stamp_begin, "time_stamp_end" : text.time_stamp_end, "person_id" : text.person_id, "language" : text.language.decode(), "confidence" : text.confidence}
+
+        if complete:
+            self._check(dll.ailiaSpeechResetTranscribeState(self._instance))
+
     def __del__(self):
         if self._instance:
             dll.ailiaSpeechDestroy(cast(self._instance, c_void_p))

{ailia_speech-1.3.0.5.dist-info → ailia_speech-1.3.1.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ailia_speech
-Version: 1.3.0.5
+Version: 1.3.1.0
 Summary: ailia AI Speech
 Home-page: https://ailia.jp/
 Author: ax Inc.
@@ -40,6 +40,10 @@ pip3 install ./
 
 ## Usage
 
+### Batch mode
+
+In batch mode, the entire audio is transcribed at once.
+
 ```python
 import ailia_speech
 
@@ -55,16 +59,63 @@ if not os.path.exists(input_file_path):
 		"https://github.com/axinc-ai/ailia-models/raw/refs/heads/master/audio_processing/whisper/demo.wa",
 		"demo.wav"
 	)
-audio_waveform, sampling_rate = librosa.load(input_file_path, mono=True)
+audio_waveform, sampling_rate = librosa.load(input_file_path, mono = True)
 
 # Infer
 speech = ailia_speech.Whisper()
-speech.initialize_model(model_path = "./models/", model_type = ailia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_SMALL)
+speech.initialize_model(model_path = "./models/", model_type = ailia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3_TURBO)
 recognized_text = speech.transcribe(audio_waveform, sampling_rate)
 for text in recognized_text:
 	print(text)
 ```
 
+### Step mode
+
+In step mode, the audio is input in chunks and transcribed sequentially.
+
+```python
+import ailia_speech
+
+import librosa
+
+import os
+import urllib.request
+
+# Load target audio
+input_file_path = "demo.wav"
+if not os.path.exists(input_file_path):
+	urllib.request.urlretrieve(
+		"https://github.com/axinc-ai/ailia-models/raw/refs/heads/master/audio_processing/whisper/demo.wa",
+		"demo.wav"
+	)
+audio_waveform, sampling_rate = librosa.load(input_file_path, mono = True)
+
+# Infer
+speech = ailia_speech.Whisper()
+speech.initialize_model(model_path = "./models/", model_type = ailia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3_TURBO)
+for i in range(0, audio_waveform.shape[0], sampling_rate):
+	complete = False
+	if i + sampling_rate >= audio_waveform.shape[0]:
+		complete = True
+	recognized_text = speech.transcribe_step(audio_waveform[i:min(audio_waveform.shape[0], i + sampling_rate)], sampling_rate, complete)
+	for text in recognized_text:
+		print(text)
+```
+
+### Available model types
+
+It is possible to select multiple models according to accuracy and speed. LARGE_V3_TURBO is the most recommended.
+
+```
+ailia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_TINY
+ilia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_BASE
+ailia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_SMALL
+ailia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_MEDIUM
+ailia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE
+ailia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3
+ailia_speech.AILIA_SPEECH_MODEL_TYPE_WHISPER_MULTILINGUAL_LARGE_V3_TURBO
+```
+
 ## API specification
 
 https://github.com/axinc-ai/ailia-sdk

{ailia_speech-1.3.0.5.dist-info → ailia_speech-1.3.1.0.dist-info}/RECORD

@@ -1,12 +1,12 @@
 ailia_speech/LICENSE_AILIA_EN.pdf,sha256=1DzVViPnw1uAS8gJ5a8uN3iZNNR5I1ItIXmezHfUpeM,70149
 ailia_speech/LICENSE_AILIA_JA.pdf,sha256=s628QN47S2bNqIfuSjm2LBf0vIluv2df6MSemn6Ksmw,174134
-ailia_speech/__init__.py,sha256=owi3WkmDlaYf0P1TJF_CVyVSW-cLMj3kp9RoVH9Rd3c,25637
+ailia_speech/__init__.py,sha256=NMFB_mXj9ed9YcbEXrjvm-82OOYRxzSBENKCuxbtz6M,27915
 ailia_speech/linux/arm64-v8a/libailia_speech.so,sha256=JAOwnBr7lbiMZmPCM99pd4vJQ08ZuXDPpq-FurrXSnE,166096
 ailia_speech/linux/x64/libailia_speech.so,sha256=WbFvA5wKTgS_Zx8ErT7WBKJbzOUexavr4nP4EkLNawQ,171360
 ailia_speech/mac/libailia_speech.dylib,sha256=-JAC40yLslAVMvfh6LhDvP3Zyt3hIT3WZc7wa9-07zU,317112
 ailia_speech/windows/x64/ailia_speech.dll,sha256=WJCOHi0Na4tdMG1RT7dA7yAoWumiGSWeW1vxUtiXDS8,126464
-ailia_speech-1.3.0.5.data/scripts/__init__.py,sha256=owi3WkmDlaYf0P1TJF_CVyVSW-cLMj3kp9RoVH9Rd3c,25637
-ailia_speech-1.3.0.5.dist-info/METADATA,sha256=webGrA8CrUKo8WSk_vIvqb_aDekuX7yzXlelkeOrP6M,1932
-ailia_speech-1.3.0.5.dist-info/WHEEL,sha256=GV9aMThwP_4oNCtvEC2ec3qUYutgWeAzklro_0m4WJQ,91
-ailia_speech-1.3.0.5.dist-info/top_level.txt,sha256=Ou9XeJ9AvdK8eutw07oosCthftD1tRYzAgNY2BrYhDc,13
-ailia_speech-1.3.0.5.dist-info/RECORD,,
+ailia_speech-1.3.1.0.data/scripts/__init__.py,sha256=NMFB_mXj9ed9YcbEXrjvm-82OOYRxzSBENKCuxbtz6M,27915
+ailia_speech-1.3.1.0.dist-info/METADATA,sha256=x5WBVorX7b9Mubn69FVh75XS2b3iDIy7ZM38F9yyGRo,3610
+ailia_speech-1.3.1.0.dist-info/WHEEL,sha256=GV9aMThwP_4oNCtvEC2ec3qUYutgWeAzklro_0m4WJQ,91
+ailia_speech-1.3.1.0.dist-info/top_level.txt,sha256=Ou9XeJ9AvdK8eutw07oosCthftD1tRYzAgNY2BrYhDc,13
+ailia_speech-1.3.1.0.dist-info/RECORD,,