Change8

v4.51.3-CSM-preview

Breaking Changes
📦 transformersView on GitHub →
1 breaking5 features🔧 6 symbols

Summary

This release introduces the Conversational Speech Model (CSM), an open-source contextual text-to-speech model capable of generating natural speech from multi-turn dialogue context.

⚠️ Breaking Changes

  • The new CSM model and related components are introduced, which requires updating the `transformers` library version.Fix: If you intend to use CSM, you must install the specific preview tag: `pip install git+https://github.com/huggingface/transformers@v4.51.3-CSM-preview`. Otherwise, no action is needed unless you rely on strict versioning, as this tag does not follow semantic versioning.

Migration Steps

  1. If you wish to use the new CSM model immediately, update your installation by running: pip install git+https://github.com/huggingface/transformers@v4.51.3-CSM-preview.
  2. If you are not using CSM, be aware that the next minor release, v4.52.0, will include this model, which might affect dependency resolution if you rely on strict minor version pinning.
  3. When using CSM, ensure you import `CsmForConditionalGeneration` from `transformers`.
  4. When preparing inputs for CSM, use the `AutoProcessor` and structure your input as a conversation list, potentially including text and audio content types, as demonstrated in the usage examples.
  5. When generating output, use `model.generate(**inputs, output_audio=True)` and process the resulting audio using `processor.save_audio()`.

✨ New Features

  • Introduction of the Conversational Speech Model (CSM), a contextual text-to-speech model.
  • Support for multi-turn dialogue context using text and spoken audio sequences.
  • Support for batched inference.
  • Support for full-graph compilation with CUDA graphs and static cache for optimized performance.
  • Integration with the Mimi codec model for discrete codebook token encoding/decoding.

Affected Symbols