app.py

import gradio as gr
import torch
import librosa
import soundfile
import nemo.collections.asr as nemo_asr
import tempfile
import os
import uuid

SAMPLE_RATE = 16000

model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained("stt_en_conformer_transducer_large")
model.change_decoding_strategy(None)
model.eval()


def process_audio_file(file):
    data, sr = librosa.load(file)

    if sr != SAMPLE_RATE:
        data = librosa.resample(data, sr, SAMPLE_RATE)

    # monochannel
    data = librosa.to_mono(data)
    return data


def transcribe(Microphone, File_Upload):
    warn_output = ""
    if (Microphone is not None) and (File_Upload is not None):
        warn_output = "WARNING: You've uploaded an audio file and used the microphone. " \
                      "The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
        file = Microphone

    elif (Microphone is None) and (File_Upload is None):
        return "ERROR: You have to either use the microphone or upload an audio file"

    elif Microphone is not None:
        file = Microphone
    else:
        file = File_Upload

    audio_data = process_audio_file(file)

    with tempfile.TemporaryDirectory() as tmpdir:
        audio_path = os.path.join(tmpdir, f'audio_{uuid.uuid4()}.wav')
        soundfile.write(audio_path, audio_data, SAMPLE_RATE)

        transcriptions = model.transcribe([audio_path])

        # if transcriptions form a tuple (from RNNT), extract just "best" hypothesis
        if type(transcriptions) == tuple and len(transcriptions) == 2:
            transcriptions = transcriptions[0]

    return warn_output + transcriptions[0]


iface = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.inputs.Audio(source="microphone", type='filepath', optional=True),
        gr.inputs.Audio(source="upload", type='filepath', optional=True),
    ],
    outputs="text",
    layout="horizontal",
    theme="huggingface",
    title="NeMo Conformer Transducer Large - English",
    description="Demo for English speech recognition using Conformer Transducers",
    allow_flagging='never',
)
iface.launch(enable_queue=True)