app.py

import os
import torch
import gradio as gr
import numpy as np
import torch
from datasets import load_dataset, Audio
from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
from speechbrain.pretrained import EncoderClassifier

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)

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

# model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
model = SpeechT5ForTextToSpeech.from_pretrained(
    "JanLilan/speecht5_finetuned_openslr-slr69-cat"
).to(device)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)

######################################################################################
################################## SPEAKER EMBEDDING #################################
######################################################################################
# we will try to translate with this voice embedding... Let's see what happen. else:
dataset = load_dataset("projecte-aina/openslr-slr69-ca-trimmed-denoised", split="train")
dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
# LOAD
spk_model_name = "speechbrain/spkrec-xvect-voxceleb"
speaker_model = EncoderClassifier.from_hparams(
    source=spk_model_name,
    run_opts={"device": device},
    savedir=os.path.join("/tmp", spk_model_name),
)

def create_speaker_embedding(waveform):
    with torch.no_grad():
        speaker_embeddings = speaker_model.encode_batch(torch.tensor(waveform))
        speaker_embeddings = torch.nn.functional.normalize(speaker_embeddings, dim=2)
        speaker_embeddings = speaker_embeddings.squeeze().cpu().numpy()
    return speaker_embeddings

# we must take one speaker embeding
checkpoint = "microsoft/speecht5_tts"
processor = SpeechT5Processor.from_pretrained(checkpoint)

# function to embedd
def prepare_dataset(example):
    audio = example["audio"]

    example = processor(
        text=example["transcription"],
        audio_target=audio["array"],
        sampling_rate=audio["sampling_rate"],
        return_attention_mask=False,
    )

    # strip off the batch dimension
    example["labels"] = example["labels"][0]

    # use SpeechBrain to obtain x-vector
    example["speaker_embeddings"] = create_speaker_embedding(audio["array"])

    return example

processed_example = prepare_dataset(dataset[2])
speaker_embeddings = torch.tensor(processed_example["speaker_embeddings"]).unsqueeze(0)

# 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": "transcribe", "language": "catalan"})
    return outputs["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 = "Demo STST - Multilingual to Català Speech"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Català. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation to català, and Microsoft's
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech fine-tuned on [projecte-aina/openslr-slr69-ca-trimmed-denoised](https://huggingface.co/datasets/projecte-aina/openslr-slr69-ca-trimmed-denoised). 

This demo can be improve updating it with [projecte-aina/tts-ca-coqui-vits-multispeaker](https://huggingface.co/projecte-aina/tts-ca-coqui-vits-multispeaker) model:

![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()