app.py

import gradio as gr
import torch
import uuid
import json
import librosa
import os
import tempfile
import soundfile as sf
import scipy.io.wavfile as wav

from transformers import pipeline, VitsModel, AutoTokenizer, set_seed
from nemo.collections.asr.models import EncDecMultiTaskModel

# Constants
SAMPLE_RATE = 16000  # Hz

# load ASR model
canary_model = EncDecMultiTaskModel.from_pretrained('nvidia/canary-1b')
decode_cfg = canary_model.cfg.decoding
decode_cfg.beam.beam_size = 1
canary_model.change_decoding_strategy(decode_cfg)

# Function to convert audio to text using ASR
def gen_text(audio_filepath, action, source_lang, target_lang):
    if audio_filepath is None:
        raise gr.Error("Please provide some input audio.")
    
    utt_id = uuid.uuid4()
    with tempfile.TemporaryDirectory() as tmpdir:
        # Convert to 16 kHz
        data, sr = librosa.load(audio_filepath, sr=None, mono=True)
        if sr != SAMPLE_RATE:
            data = librosa.resample(data, orig_sr=sr, target_sr=SAMPLE_RATE)
        converted_audio_filepath = os.path.join(tmpdir, f"{utt_id}.wav")
        sf.write(converted_audio_filepath, data, SAMPLE_RATE)

        # Transcribe audio
        duration = len(data) / SAMPLE_RATE
        manifest_data = {
            "audio_filepath": converted_audio_filepath,
            "taskname": action,
            "source_lang": source_lang,
            "target_lang": source_lang if action=="asr" else target_lang,
            "pnc": "no",
            "answer": "predict",
            "duration": str(duration),
        }
        manifest_filepath = os.path.join(tmpdir, f"{utt_id}.json")
        with open(manifest_filepath, 'w') as fout:
            fout.write(json.dumps(manifest_data))

        predicted_text = canary_model.transcribe(manifest_filepath)[0]
        # if duration < 40:
        #     predicted_text = canary_model.transcribe(manifest_filepath)[0]
        # else:
        #     predicted_text = get_buffered_pred_feat_multitaskAED(
        #         frame_asr,
        #         canary_model.cfg.preprocessor,
        #         model_stride_in_secs,
        #         canary_model.device,
        #         manifest=manifest_filepath,
        #     )[0].text
    
    return predicted_text

# Function to convert text to speech using TTS
def gen_speech(text, lang):
    set_seed(555)  # Make it deterministic
    match lang:
        case "en":
             model = "facebook/mms-tts-eng"
        case "fr":
             model = "facebook/mms-tts-fra"
        case "de":
             model = "facebook/mms-tts-deu"
        case "es":
             model = "facebook/mms-tts-spa"
        case _:
            model = "facebook/mms-tts-eng"
    
    # load TTS model
    tts_model = VitsModel.from_pretrained(model)
    tts_tokenizer = AutoTokenizer.from_pretrained(model)

    input_text = tts_tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        outputs = tts_model(**input_text)
    waveform_np = outputs.waveform[0].cpu().numpy()
    output_file = f"{str(uuid.uuid4())}.wav"
    wav.write(output_file, rate=tts_model.config.sampling_rate, data=waveform_np)
    return output_file

# Root function for Gradio interface
def start_process(audio_filepath, source_lang, target_lang):
    transcription = gen_text(audio_filepath, "asr", source_lang, target_lang)
    print("Done transcribing")
    translation = gen_text(audio_filepath, "s2t_translation", source_lang, target_lang) 
    print("Done translation")
    audio_output_filepath = gen_speech(translation, target_lang)
    print("Done speaking")   
    return transcription, translation, audio_output_filepath

    
# Create Gradio interface
playground = gr.Blocks()

with playground:

    with gr.Row():
        gr.Markdown("""
                    ## Your AI Translate Assistant
                    ### Gets input audio from user, transcribe and translate it. Convert back to speech.
                    - category: Automatic Speech Recognition, model: [nvidia/canary-1b](https://huggingface.co/nvidia/canary-1b)
                    - category: Text-to-Speech, model: [facebook/mms-tts-eng](https://huggingface.co/facebook/mms-tts-eng)
                    """)
        
    with gr.Row():
        with gr.Column():
            source_lang = gr.Dropdown(
                choices=["en", "de", "es", "fr"], value="en", label="Source Language"
            )
        with gr.Column():
            target_lang = gr.Dropdown(
                choices=["en", "de", "es", "fr"], value="fr", label="Target Language"
            )            

    with gr.Row():
        with gr.Column():
            input_audio = gr.Audio(sources=["microphone"], type="filepath", label="Input Audio")            
        with gr.Column():
            translated_speech = gr.Audio(type="filepath", label="Generated Speech")            

    with gr.Row():
        with gr.Column():            
            transcipted_text = gr.Textbox(label="Transcription")
        with gr.Column():            
            translated_text = gr.Textbox(label="Translation")

    with gr.Row():
        with gr.Column():
            submit_button = gr.Button(value="Start Process", variant="primary")
        with gr.Column():
            clear_button = gr.ClearButton(components=[input_audio, source_lang, target_lang, transcipted_text, translated_text, translated_speech], value="Clear")    

    with gr.Row():
        gr.Examples(
            examples=[
                ["sample_en.wav","en","fr"],
                ["sample_fr.wav","fr","de"],
                ["sample_de.wav","de","es"],
                ["sample_es.wav","es","en"]
            ], 
            inputs=[input_audio, source_lang, target_lang], 
            outputs=[transcipted_text, translated_text, translated_speech], 
            run_on_click=True, cache_examples=True, fn=start_process
        )
            
    submit_button.click(start_process, inputs=[input_audio, source_lang, target_lang], outputs=[transcipted_text, translated_text, translated_speech])

playground.launch()