# Licensing Details for the Speech-to-Speech Models: Repository Code vs. Model Weights

> Understand the licensing for speech-to-speech models. Explore the Apache 2.0 repo license and individual model licenses for Parler TTS, Melo TTS, and more.

- Repository: [Hugging Face/speech-to-speech](https://github.com/huggingface/speech-to-speech)
- Tags: licensing-details
- Published: 2026-08-01

---

**The huggingface/speech-to-speech repository is released under Apache License 2.0, but each speech-to-speech model (Parler TTS, Melo TTS, Moonshine STT, etc.) carries its own separate license specified on its Hugging Face Hub model card.**

The `huggingface/speech-to-speech` repository provides a unified inference framework for end-to-end speech-to-speech translation and generation. Understanding the licensing details for the speech-to-speech models is essential for legal compliance, as the codebase and the downloadable model weights operate under distinctly different licensing frameworks.

## Repository License vs. Model Licenses

The `speech-to-speech` project employs a dual-layer licensing structure that separates the **inference code** from the **model weights**. The repository's source code—including the inference pipeline, utility scripts, and demo implementations—is uniformly licensed under Apache License 2.0. However, the actual speech-to-speech models that the library loads at runtime are not bundled with the source code; instead, they are downloaded separately from the Hugging Face Hub, where each model maintains its own license terms.

## Apache 2.0 Coverage for the Codebase

All source code within the repository falls under the permissive Apache License 2.0. You can verify this in the repository root:

- The **`LICENSE`** file contains the full Apache 2.0 legal text
- The **[`pyproject.toml`](https://github.com/huggingface/speech-to-speech/blob/main/pyproject.toml)** declares `license = "Apache-2.0"` in the project metadata
- The **`src/speech_to_speech/`** directory, which contains the core library implementation including `SpeechToSpeechPipeline` and processor handlers, is entirely Apache 2.0

This license grants you broad rights to use, modify, distribute, and sublicense the code, including for commercial applications, provided you include the original copyright notice and disclaimer.

## Per-Model Licensing on the Hugging Face Hub

When you instantiate a pipeline with specific models, you are loading weights distributed under separate licenses. Common speech-to-speech models and their typical license structures include:

- **Parler TTS** – Usually Apache 2.0, but always verify the model card
- **Melo TTS** – Typically permissive open-source licenses
- **Moonshine STT** – License varies by checkpoint version

Typical license categories you will encounter include:

- **Apache-2.0** – Many open-source TTS/STT models follow the same permissive terms as the library
- **MIT** – Research prototypes often use this permissive license
- **CC-BY-4.0** or **CC-BY-SA-4.0** – Some audio-generation models require attribution or share-alike compliance

The repository does not redistribute these licenses. Files like [`archive/TTS/parler_handler.py`](https://github.com/huggingface/speech-to-speech/blob/main/archive/TTS/parler_handler.py) contain the code to load and run Parler TTS, but the actual license terms for the Parler TTS weights reside on its Hugging Face Hub model card.

## Checking License Compliance in Practice

Always inspect the model card before downloading weights. The following code demonstrates loading models while emphasizing the need to verify individual license terms:

```python
from speech_to_speech import SpeechToSpeechPipeline
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq

# Load processor and model - check the Hugging Face Hub page for license specifics

# Parler TTS typically uses Apache-2.0, but verify at huggingface.co/parler-tts/parler-tts

processor = AutoProcessor.from_pretrained("parler-tts/parler-tts")
tts_model = AutoModelForSpeechSeq2Seq.from_pretrained("parler-tts/parler-tts")

# Initialize pipeline (pipeline code is Apache-2.0)

pipeline = SpeechToSpeechPipeline(
    tts_processor=processor,
    tts_model=tts_model,
    # Additional components (STT, LLM) would be loaded here with their own licenses

)

# Generate audio - usage rights depend on the TTS model's license, not the pipeline's

output_audio = pipeline.run(text="Hello, world!")

```

Users bear full responsibility for complying with each model's license when redistributing or commercializing generated audio.

## Summary

- The **huggingface/speech-to-speech codebase** (including `src/speech_to_speech/` and `SpeechToSpeechPipeline`) is **Apache License 2.0**
- **Model weights** (Parler TTS, Melo TTS, Moonshine STT, etc.) carry **individual licenses** specified on their Hugging Face Hub model cards
- **License types vary**—common options include Apache-2.0, MIT, and Creative Commons variants
- Users must **verify model cards** before commercial use or redistribution of generated content

## Frequently Asked Questions

### Is the huggingface/speech-to-speech codebase free for commercial use?

Yes. The repository's source code, including the inference pipeline and handler implementations like [`archive/TTS/parler_handler.py`](https://github.com/huggingface/speech-to-speech/blob/main/archive/TTS/parler_handler.py), is released under Apache License 2.0. This permits commercial use, modification, and distribution, provided you include the required attribution and disclaimer.

### Do all speech-to-speech models use the same license as the repository?

No. While the code is uniformly Apache 2.0, the models you load (such as Parler TTS or Moonshine STT) are distributed under their own licenses via the Hugging Face Hub. These may differ from the repository's license and can include Apache 2.0, MIT, or Creative Commons licenses.

### Where can I find the license for a specific model like Parler TTS?

Each model's license is displayed on its Hugging Face Hub model card. When you call `AutoProcessor.from_pretrained()` or `AutoModelForSpeechSeq2Seq.from_pretrained()`, the library downloads weights from the Hub, where the license metadata is explicitly stated. The `speech-to-speech` repository does not mirror or override these terms.

### Can I redistribute audio generated using these models?

Redistribution rights depend entirely on the specific licenses of the models used to generate the audio. While the Apache 2.0 license governs the code that produced the audio, the output itself may be subject to the terms of the TTS model (such as attribution requirements under CC-BY or share-alike obligations under CC-BY-SA). Always consult the model card for the specific checkpoint you are using.