Skip to content
Equinox Paper EP markEPequinoxpaper

Weighing the forces shaping machine intelligence.

Language models

Phonon TTS delivers offline voice synthesis with Apache 2.0 licensing for commercial use

Hugging Face’s small on-device speech model runs without servers, supports five languages and allows free commercial deployment of its English variant.

Phonon TTS delivers offline voice synthesis with Apache 2.0 licensing for commercial use
Depth
31mPhonon 31m is licensed under Apache 2.0, permitting free commercial use for English speech generation · reported
1 outlet · newest source 168h ago

Key points

  • Phonon 31m is licensed under Apache 2.0, permitting free commercial use for English speech generation.
  • The model operates entirely on-device, requiring no network connection or server round-trips for audio synthesis.
  • Two model sizes, approximately 65 MB and 75 MB, support English, French, German, Spanish and Portuguese.

Hugging Face published a blog post detailing the capabilities and licensing of Phonon, its small text-to-speech model designed for on-device execution. The release highlights that the system generates speech locally, keeping user text private and eliminating dependency on external servers or network connectivity.The model lets users type text and hear it spoken aloud on their own device, without sending data to the internet.Phonon employs flow matching within the Pocket TTS lineage to achieve competitive quality in sub-100 MB models.

Evidence: confirmedA Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.

aifleet-huggingface

The evidence

According to the Hugging Face blog post, Phonon was initially announced in April and underwent six months of development with partners and early users. This feedback period resulted in smaller model sizes and expanded language support, including English, French, German, Spanish and Portuguese. The post states that Phonon 31m, which speaks English, is available under an Apache 2.0 license for free commercial use. Larger models covering all five languages are licensed under CC BY-NC for non-commercial projects, with commercial pricing based on user count rather than generation volume.

Evidence: reportedA Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.

aifleet-huggingface

Evidence: reportedA Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.

aifleet-huggingface

The decision

The documentation indicates that Phonon is based on the Pocket TTS lineage and uses flow matching to generate audio. This technique differs from other text-to-speech systems and enabled high quality in a small model, though it introduced numerical instability during training that required specific adjustments. The blog post notes that the models handle complex inputs such as phone numbers, dates and codes, which often challenge other systems.

What the evidence means

The source material describes a trade-off between model size and linguistic coverage. The smaller 65 MB model supports only English and carries the permissive Apache 2.0 license. The larger 75 MB model supports five languages but is restricted to non-commercial use by default. Developers seeking multi-language commercial deployment must negotiate pricing based on their user base, as stated in the blog post.

What builders can do with it

Developers can integrate Phonon into applications running on phones, laptops, single-board computers and browser tabs. The Hugging Face team built a Transformers integration, allowing immediate use in that library. The blog post lists demos including an offline coffee brew guide on Apple Watch, a fitness counter on iPhone without signal, and a browser-based dinosaur game with voice narration. Third-party conversions for ExecuTorch and Node.js via Micdrop are also documented.

Why it matters

  • The Apache 2.0 license for the English model removes legal barriers for commercial products requiring on-device speech synthesis.
  • On-device operation ensures user text data remains local, addressing privacy concerns associated with cloud-based voice AI.
  • Support for complex entities like dates and codes expands the model's utility for practical application interfaces.
  • The small file size of approximately 65 MB enables deployment on resource-constrained devices such as smartwatches.

Topics: text-to-speech, Hugging Face, on-device AI, open-source models, speech synthesis

Updated

Then / now / next

Then

Most voice AI systems required cloud servers, sending every sentence for processing and returning audio.

Now

Phonon enables high-quality speech generation directly on user devices without network dependency.

Next

Monitor the adoption of third-party integrations like ExecuTorch and Node.js conversions mentioned in the post.

Story timeline

  1. Then Most voice AI systems required cloud servers, sending every sentence for processing and returning audio.
  2. 05 Nov 2026 A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.
  3. Now Phonon enables high-quality speech generation directly on user devices without network dependency.
  4. Next Monitor the adoption of third-party integrations like ExecuTorch and Node.js conversions mentioned in the post.

For readers

Developers

  • Use the Transformers integration to implement Phonon in Python applications immediately.
  • Leverage ExecuTorch or Micdrop conversions for browser GPU or Node.js environments.

Policy

  • On-device processing keeps user text local, reducing data transmission risks.
  • Licensing varies by model size, with commercial options for English only.

Business

  • Commercial multi-language use requires pricing based on user count, not generation volume.
  • The Apache 2.0 English model allows free commercial integration without licensing fees.

Ask the article

Local keyword search of this article; no generated answers.

Which of these is confirmed?

Check the answer

Hugging Face published a blog post detailing the capabilities and licensing of Phonon, its small text-to-speech model designed for on-device execution.

The matched primary source and announcement wording support the confirmed label.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Terms in this story

Flow matching

A modeling technique used by Phonon to generate audio, differing from other TTS systems.

Apache 2.0

A permissive open-source license allowing free commercial use of the Phonon 31m model.

CC BY-NC

A Creative Commons license restricting use to non-commercial projects, applied to larger Phonon models.

On-device

Running software locally on the user's hardware without relying on external servers.

Questions readers ask

Can I use Phonon in a commercial product?

Yes, for the English-only Phonon 31m model under Apache 2.0, or for multi-language models if you agree to user-based pricing.

Does Phonon require an internet connection?

No, the model runs entirely on the local device without needing a server or network connection.

Which languages does Phonon support?

The 75 MB model supports English, French, German, Spanish and Portuguese; the 65 MB model supports only English.

Connected stories

This storyTether AI’s BotPsy offers sub-billion-parameter agentic control for on-device workflowsTether AI’s BotPsy offers sub-biApple shifts AI spend from data centres to on-device R&D amid US election uncertaintyApple shifts AI spend from data Swift devs swap Apple speech for local Parakeet and Qwen3-ASR modelsSwift devs swap Apple speech for

Reading receipt

  • Phonon 31m is licensed under Apache 2.0, permitting free commercial use for English speech generation.
  • The model operates entirely on-device, requiring no network connection or server round-trips for audio synthesis.
  • Two model sizes, approximately 65 MB and 75 MB, support English, French, German, Spanish and Portuguese.
Time spent: 0:00 · Depth: Full

Sources

  1. Hugging Face

Related stories