app.py

import gradio as gr
import numpy as np
import torch
from datasets import load_dataset

from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline

from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer
# from tokenization_small100 import SMALL100Tokenizer


device = "cuda:0" if torch.cuda.is_available() else "cpu"

# load speech translation checkpoint


asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device, generate_kwargs = {"task": "translate"})

m2m100_en_sw = pipeline('translation', 'facebook/m2m100_418M', src_lang='en', tgt_lang="sw")

# load text-to-speech checkpoint and speaker embeddings
# processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
# debug build
processor = SpeechT5Processor.from_pretrained("samuelleecong/speecht5_finetuned_swahili")


model = SpeechT5ForTextToSpeech.from_pretrained("samuelleecong/speecht5_finetuned_swahili").to(device)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)

embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)


def translate(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
    output_translated = m2m100_en_sw(outputs["text"])
    # encoded_text = tokenizer(outputs["text"], return_tensors="pt")
    # generated_tokens = model.generate(**encoded_text)
    # output_translated = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
    # return outputs["text"]
    return output_translated[0]["translation_text"]

def synthesise(text):
    inputs = processor(text=text, return_tensors="pt")
    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
    return speech.cpu()


def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    synthesised_speech = synthesise(translated_text)
    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
    return 16000, synthesised_speech


title = "Cascaded STST"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Swahili. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech transcription, Meta's [M2M100](https://huggingface.co/facebook/m2m100_418M) for translation, and Microsoft's
[SpeechT5 TTS](https://huggingface.co/samuelleecong/speecht5_finetuned_swahili) that I fine-tuned model for Swahili text-to-speech:

![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
"""

demo = gr.Blocks()

mic_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="microphone", type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    title=title,
    description=description,
)

file_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="upload", type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    examples=[["./example.wav"]],
    title=title,
    description=description,
)

with demo:
    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])

demo.launch()