app.py

import gradio as gr
import torch
import time
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("nvidia/stt_en_conformer_transducer_xlarge")
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, orig_sr=sr, target_sr=SAMPLE_RATE)

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


def transcribe(audio, state=""):
    # Grant additional context
    # time.sleep(1)

    if state is None:
        state = ""

    audio_data = process_audio_file(audio)

    with tempfile.TemporaryDirectory() as tmpdir:
        # Filepath transcribe
        audio_path = os.path.join(tmpdir, f'audio_{uuid.uuid4()}.wav')
        soundfile.write(audio_path, audio_data, SAMPLE_RATE)
        transcriptions = model.transcribe([audio_path])
        
        # Direct transcribe
        # transcriptions = model.transcribe([audio])

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

        transcriptions = transcriptions[0]

    state = state + transcriptions + " "
    return state, state


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