app.py

import gradio as gr
import torch

import nemo.collections.asr as nemo_asr

SAMPLE_RATE = 16000
TITLE = "NeMo ASR Inference on Hugging Face"
DESCRIPTION = "Demo of all languages supported by NeMo ASR"
DEFAULT_EN_MODEL = "nvidia/stt_en_conformer_transducer_xlarge"

MARKDOWN = f"""
# {TITLE}

## {DESCRIPTION}
"""

CSS = """
p.big {
  font-size: 20px;
}
"""

ARTICLE = """
<br><br>
<p class='big' style='text-align: center'>
    <a href='https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/asr/intro.html' target='_blank'>NeMo ASR</a> 
    | 
    <a href='https://github.com/NVIDIA/NeMo#nvidia-nemo' target='_blank'>Github Repo</a>
</p>
"""

SUPPORTED_LANGUAGES = set([])
SUPPORTED_MODEL_NAMES = set([])

# HF models
hf_filter = nemo_asr.models.ASRModel.get_hf_model_filter()
hf_filter.task = "automatic-speech-recognition"

hf_infos = nemo_asr.models.ASRModel.search_huggingface_models(model_filter=hf_filter)
for info in hf_infos:
    lang_id = info.modelId.split("_")[1]  # obtains lang id as str
    SUPPORTED_LANGUAGES.add(lang_id)
    SUPPORTED_MODEL_NAMES.add(info.modelId)

SUPPORTED_MODEL_NAMES = sorted(list(SUPPORTED_MODEL_NAMES))

model_dict = {model_name: gr.Interface.load(f'models/{model_name}') for model_name in SUPPORTED_MODEL_NAMES}

SUPPORTED_LANG_MODEL_DICT = {}
for lang in SUPPORTED_LANGUAGES:
    for model_id in SUPPORTED_MODEL_NAMES:
        if lang in model_id:
            # create new lang in dict
            if lang not in SUPPORTED_LANG_MODEL_DICT:
                SUPPORTED_LANG_MODEL_DICT[lang] = [model_id]
            else:
                SUPPORTED_LANG_MODEL_DICT[lang].append(model_id)

# Sort model names
for lang in SUPPORTED_LANG_MODEL_DICT.keys():
    model_ids = SUPPORTED_LANG_MODEL_DICT[lang]
    model_ids = sorted(model_ids)
    SUPPORTED_LANG_MODEL_DICT[lang] = model_ids


def transcribe(microphone, audio_file, model_name):
    model = model_dict[model_name]

    warn_output = ""
    if (microphone is not None) and (audio_file 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"
        )
        audio_data = microphone

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

    elif microphone is not None:
        audio_data = microphone
    else:
        audio_data = audio_file

    try:
        # Use HF API for transcription
        transcriptions = model(audio_data)

    except Exception as e:
        transcriptions = ""
        warn_output = warn_output + "\n\n"
        warn_output += (
            f"The model `{model_name}` is currently loading and cannot be used "
            f"for transcription.\n"
            f"Please try another model or wait a few minutes."
        )

    return warn_output + transcriptions


demo = gr.Blocks(title=TITLE, css=CSS)

with demo:
    header = gr.Markdown(MARKDOWN)

    with gr.Row() as row:
        file_upload = gr.components.Audio(source="upload", type='filepath', label='Upload File')
        microphone = gr.components.Audio(source="microphone", type='filepath', label='Microphone')

    lang_selector = gr.components.Dropdown(
        choices=sorted(list(SUPPORTED_LANGUAGES)), value="en", type="value", label="Languages", interactive=True,
    )
    models_in_lang = gr.components.Dropdown(
        choices=sorted(list(SUPPORTED_LANG_MODEL_DICT["en"])),
        value=DEFAULT_EN_MODEL,
        label="Models",
        interactive=True,
    )

    def update_models_with_lang(lang):
        models_names = sorted(list(SUPPORTED_LANG_MODEL_DICT[lang]))
        default = models_names[0]

        if lang == 'en':
            default = DEFAULT_EN_MODEL
        return models_in_lang.update(choices=models_names, value=default)

    lang_selector.change(update_models_with_lang, inputs=[lang_selector], outputs=[models_in_lang])

    transcript = gr.components.Label(label='Transcript')

    run = gr.components.Button('Transcribe')
    run.click(transcribe, inputs=[microphone, file_upload, models_in_lang], outputs=[transcript])

    gr.components.HTML(ARTICLE)

demo.queue(concurrency_count=1)
demo.launch()