asr_kaldi_gradio.py

import json
import gradio as gr
import numpy as np

from vosk import KaldiRecognizer, Model

model = Model(model_path="/home/gns/vosk-api/model/")


def reformat_freq(sr, y):
    if sr not in (
        48000,
        16000,
        8000
    ):
        raise ValueError("Unsupported rate", sr)
    if sr == 48000:
        y = (
            ((y / max(np.max(y), 1)) * 32767)
            .reshape((-1, 6))
            .mean(axis=1)
            .astype("int16").tobytes()
        )
        sr = 8000
    elif sr == 16000:
        y = (
            ((y / max(np.max(y), 1)) * 32767)
            .reshape((-1, 2))
            .mean(axis=1)
            .astype("int16")
        )
        sr = 8000
    return sr, y

def transcribe(data, state):
    sample_rate, audio_data = data
    sr, y = reformat_freq(sample_rate, audio_data)
    print('new sample rate: '+str(sr))

    if state is None:
        rec = KaldiRecognizer(model, 8000)
        result = []
    else:
        rec, result = state

    if rec.AcceptWaveform(y):
        text_result = json.loads(rec.Result())["text"]
        if text_result != "":
            result.append(text_result)
        partial_result = ""
    else:
        partial_result = json.loads(rec.PartialResult())["partial"] + " "

    print("\n".join(result) + "\n" + partial_result, (rec, result))
    return "\n".join(result) + "\n" + partial_result, (rec, result)

gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(source="microphone", type="numpy", streaming=True),
        "state"
    ],
    outputs=[
        "textbox",
        "state"
    ],
    live=True).launch(share=True)