update

main.py

@@ -1,27 +1,57 @@
 from flask import Flask, request, jsonify
 from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
-from transformers import pipeline
 import numpy as np
-import json
+import torch
+
 app = Flask(__name__)
 model = AutoModelForSpeechSeq2Seq.from_pretrained("GitNazarov/whisper-small-pt-3-uz")
 processor = AutoProcessor.from_pretrained("GitNazarov/whisper-small-pt-3-uz")
 
+
+USE_ONNX = False # change this to True if you want to test onnx model
+silero_vad_path = 'snakers4/silero-vad'
+vad_model, vad_utils = torch.hub.load(silero_vad_path,
+                              model='silero_vad',
+                              force_reload=True,
+                              onnx=USE_ONNX)
+
+(get_speech_timestamps,
+save_audio,
+read_audio,
+VADIterator,
+collect_chunks) = vad_utils
+STT_SAMPLE_RATE = 16000
+
+
+def int2float(sound):
+    abs_max = np.abs(sound).max()
+    sound = sound.astype('float32')
+    if abs_max > 0:
+        sound *= 1/32768
+    sound = sound.squeeze()  # depends on the use case
+    return sound
+
 @app.route('/', methods=['GET'])
 def index():
     return jsonify({"message": "Welcome to whisper uz!"})
 
 @app.route('/transcribe', methods=['POST'])
 def transcribe():
-    data = request.json['data']
-    data = json.loads(data)
-    tensor_data = np.array(data)
-    inputs = processor(tensor_data, return_tensors="pt", sampling_rate=16000, max_new_tokens=100)
-    input_features = inputs.input_features
-    generated_ids = model.generate(inputs=input_features)
-
-    transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)
-    transcription = ''.join(transcription)
+    data_frames = request.data
+    audio_data = np.frombuffer(data_frames, dtype=np.int16)
+    audio_float = int2float(audio_data)
+    final_data = torch.from_numpy(audio_float)
+    sp_timestamps = get_speech_timestamps(final_data, vad_model, sampling_rate=STT_SAMPLE_RATE)
+    try:
+        final_audio_data = collect_chunks(sp_timestamps, final_data)
+        inputs = processor(final_audio_data, return_tensors="pt", sampling_rate=16000, max_new_tokens=100)
+        input_features = inputs.input_features
+        generated_ids = model.generate(inputs=input_features)
+
+        transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        transcription = ''.join(transcription)
+    except Exception as e:
+        transcription = str(e)
     return str(transcription), {'Content-Type': 'application/json'}
 
 if __name__ == '__main__':