Fine-tuning text-to-speech models

Multi-Modal Models with Hugging Face

James Chapman

Curriculum Manager, DataCamp

Purpose of fine-tuning text-to-speech

Learn sounds in new languages and dialects
Apply to new contexts

E.g., large English language for general pretraining model ⇏ realistic Italian speech

cartoon english italian speech bubbles

Purpose of fine-tuning text-to-speech

Diagram of how a speech embedding fits in a speech generation pipeline

Speaker embedding + text-to-speech model features → generative model
New speaker embedding are insufficient on their own without fine-tuning

Preparing an audio dataset

VoxPopuli dataset: transcribed speech data for 18 languages from EU parliament

from datasets import load_dataset
dataset = load_dataset("facebook/voxpopuli", "it", split="train", 
                       trust_remote_code=True)
print(dataset.features)

['audio', 'raw_text', 'normalized_text', 'gender', 'speaker_id', ... ]

Need to preprocess the audio + add speech embeddings:

speaker_model = EncoderClassifier.from_hparams(source="speechbrain/spkrec-xvect-voxceleb", 
                                       savedir="pretrained_models/spkrec-xvect-voxceleb")

Audio preprocessing

from transformers import SpeechT5Processor
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")


def prepare_dataset(example):
    audio = example["audio"]
    example = processor(text=example["normalized_text"], audio_target=audio["array"], 
        sampling_rate=audio["sampling_rate"], return_attention_mask=False)

    example["labels"] = example["labels"][0]

    with torch.no_grad():
        speaker_embeddings = speaker_model.encode_batch(torch.tensor(audio["array"]))
        speaker_embeddings = torch.nn.functional.normalize(speaker_embeddings, dim=2)
        example["speaker_embeddings"] = speaker_embeddings.squeeze().cpu().numpy()
    return example


dataset = dataset.map(prepare_dataset)

Training arguments

from transformers import Seq2SeqTrainingArguments


training_args = Seq2SeqTrainingArguments( 
        per_device_train_batch_size=4, 

        gradient_accumulation_steps=8,

        learning_rate=1e-5,

        warmup_steps=500,

        label_names=["labels"],

        data_collator=data_collator
)

Putting it all together

model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")

Trainer:

trainer = Seq2SeqTrainer(args=training_args, model=model,
    train_dataset=dataset["train"], eval_dataset=dataset["test"],
                         tokenizer=processor)

Run the training: trainer.train()

Using the new model

text = "se sono italiano posso cantare l'opera lirica"


speaker_embedding = torch.tensor(dataset[5]["speaker_embeddings"]).unsqueeze(0)

inputs = processor(text=text, return_tensors="pt")

speech = model.generate_speech(inputs["input_ids"],
                               speaker_embedding,
                               vocoder=vocoder)


make_spectrogram(speech)

Using the new model

Italian speech spectrogram

Let's practice!

Multi-Modal Models with Hugging Face