Fine-tuning a text-to-speech model

You will be working with the VCTK Corpus, which includes around 44-hours of speech data uttered by English speakers with various accents, to fine-tune a text-to-speech model to replicate regional accents.

The dataset has already been loaded and preprocessed, and the SpeechT5ForTextToSpeech module has been loaded, as have the Seq2SeqTrainingArguments and Seq2SeqTrainer modules. A data collator (data_collator) has been predefined.

Please do not call the .train() method on the trainer config, as this code will time out in this environment.

This exercise is part of the course

Multi-Modal Models with Hugging Face

Exercise instructions

Load the microsoft/speecht5_tts pretrained model using SpeechT5ForTextToSpeech.
Create an instance of Seq2SeqTrainingArguments with: gradient_accumulation_steps set to 8, learning_rate set to 0.00001, warmup_steps set to 500, and max_steps set to 4000.
Configure the trainer with the new training arguments, and the model, data, and processor provided.

Hands-on interactive exercise

Have a go at this exercise by completing this sample code.

# Load the text-to-speech pretrained model
model = ____.____(____)

# Configure the required training arguments
training_args = ____(output_dir="speecht5_finetuned_vctk_test",
    gradient_accumulation_steps=____, learning_rate=____, warmup_steps=____, max_steps=4000, label_names=["labels"],
    push_to_hub=False)

# Configure the trainer
trainer = ____(args=training_args, model=model, data_collator=data_collator,
    train_dataset=dataset["train"], eval_dataset=dataset["test"], tokenizer=processor)

Edit and Run Code

This exercise is part of the course

Multi-Modal Models with Hugging Face

IntermediateSkill Level

4.9+

Start Course for Free

Navigate the Hugging Face model hub, transform raw text, audio, and visual data into AI-friendly formats. Learn how to find the latest most popular models for tasks such as text generation and harness the power of pre-built pipelines.

Exercise 1: Hugging Face model navigation Exercise 2: How many models!?Exercise 3: Finding the most popular text-to-image model Exercise 4: Preprocessing different modalities Exercise 5: Text tokenizing Exercise 6: Image preprocessing Exercise 7: Audio preprocessing Exercise 8: Pipeline tasks and evaluations Exercise 9: Pipeline caption generation Exercise 10: Passing keyword arguments Exercise 11: Model evaluation on a custom dataset

Learn to master individual modalities with state-of-the-art models. Dive into computer vision for image classification and segmentation, explore speech recognition and text-to-speech synthesis, and learn effective fine-tuning techniques. Build practical skills with pre-trained models from Hugging Face's transformers library.

Exercise 1: Computer vision Exercise 2: Image classification Exercise 3: Object detection Exercise 4: Image background removal Exercise 5: Fine-tuning computer vision models Exercise 6: CV fine-tuning: dataset prep Exercise 7: CV fine-tuning: model classes Exercise 8: CV fine-tuning: trainer configuration Exercise 9: Speech recognition and audio generation Exercise 10: Automatic speech recognition Exercise 11: Creating speech embeddings Exercise 12: Audio denoising Exercise 13: Fine-tuning text-to-speech models Exercise 14: Fine-tuning a text-to-speech model

Current Exercise

Exercise 15: Generating new speech

Learn to fuse visual, textual, and audio information for richer AI applications. Master techniques like CLIP for zero-shot classification, build sentiment analyzers that see and read, and create emotion detectors that combine facial expressions with voice. Take your AI models beyond single-modality thinking.

Exercise 1: Zero-shot image classification Exercise 2: Zero-shot learning with CLIP Exercise 3: Automated caption quality assessment Exercise 4: Multi-modal sentiment analysis Exercise 5: Prompting Vision Language Models (VLMs)Exercise 6: Multi-modal sentiment classification with Qwen Exercise 7: Zero-shot video classification Exercise 8: Video audio splitting Exercise 9: Video sentiment analysis with CLIP CLAP

Transform ideas into reality! Master cutting-edge AI techniques to generate and manipulate visual content using text prompts. Create stunning images, edit photos intelligently, and build powerful question-answering systems for images and documents. Turn your creative vision into digital reality with multi-modal AI.

Exercise 1: Visual question-answering (VQA)Exercise 2: VQA with Vision Language Transformers (ViLTs)Exercise 3: Document VQA with LayoutLM Exercise 4: Image editing with diffusion models Exercise 5: Custom image editing Exercise 6: Image inpainting Exercise 7: Video generation Exercise 8: Build a video!Exercise 9: Assessing video generation performance Exercise 10: Congratulations!