KaldiRecognizer + ffmpeg

import asyncio
import websockets
from pydub import AudioSegment
import io
from vosk import Model, KaldiRecognizer
import subprocess


desired_bitrate = "1411k"
desired_samplerate = "44100"


vosk_model = Model("./model")
recognizer = KaldiRecognizer(vosk_model, 44100)

recognizer.SetWords(True)  # Enable word-level timing information in the recognition results
recognizer.SetMaxAlternatives(10)  # Adjust the maximum number of alternatives


def save_audio_to_wav(audio_data, filename):
    # Convert WebM audio data to WAV
    audio_segment = AudioSegment.from_file(io.BytesIO(audio_data), format="webm")
    audio_segment.export(filename, format="wav")

    # Run ffmpeg command for downsampling and controlling bitrate
    ffmpeg_command = [
        "ffmpeg",
        "-y",  # Automatically overwrite files
        "-i", filename,
        "-ar", desired_samplerate, # Set the sample rate to 44100 Hz
        "-b:a", desired_bitrate,   # Set the bitrate
        filename.replace(".webm", ".wav")
    ]

    subprocess.run(ffmpeg_command)


async def audio_handler(websocket, path):
    print("WebSocket connection established")

    try:
        async for message in websocket:
            save_audio_to_wav(message, "temp.webm")  # Save WebM audio to a temporary WAV file
            recognizer.AcceptWaveform(open("temp.wav", "rb").read())

            result = recognizer.Result()
            print("Result:", result)


        # Get the final recognition result
        result = recognizer.FinalResult()
        print("Final Result:", result)


    except websockets.ConnectionClosed:
        print("WebSocket connection closed")


if __name__ == "__main__":
    start_server = websockets.serve(audio_handler, "localhost", 8765)

    asyncio.get_event_loop().run_until_complete(start_server)
    print("WebSocket server started")

    try:
        asyncio.get_event_loop().run_forever()
    except KeyboardInterrupt:
        pass
    finally:
        print("WebSocket server stopped")