Phonon TTS Brings High-Quality On-Device Speech to Local AI
Gradium's open-weight Phonon TTS models run entirely offline on phones, laptops and browsers, delivering top benchmark scores for intelligibility and speed while keeping user…
- Phonon runs offline on diverse hardware including phones, single-board computers and browsers with no network required.
- It tops CPU streaming benchmarks for time-to-first-audio and ranks first in German and Spanish intelligibility on the new Hugging Face Open TTS Leaderboard.
- Weights are available under Apache 2.0 for the English 31m model, enabling free commercial use, while the multilingual 40m uses a non-commercial license.
Offline Voice Generation on Everyday Devices
Gradium released Phonon TTS two weeks ago, a family of small text-to-speech models that generate speech directly on user devices without servers or internet. The models, around 65-75 MB, support English, French, German, Spanish and Portuguese with more than 100 built-in voices.
Users can create custom voices through cloning or text description in Gradium Studio. The system handles practical elements such as phone numbers, dates and codes that often challenge other TTS tools. Demos include offline guidance on an Apple Watch, workout coaching on an iPhone without signal, and voice acting in games and browser experiences.
Inference runs through a dedicated ptts framework supporting over 10 backends including CPU, GPU, NPU and WebAssembly, or via Hugging Face Transformers for batched GPU workloads. The 31m English model is licensed under Apache 2.0 for unrestricted commercial use.
Benchmark Performance and Technical Approach
On the new Hugging Face Open TTS Leaderboard, Phonon ranks first in intelligibility for German and Spanish and in the top three across evaluated languages, despite being up to 100 times smaller than competing models. It leads the CPU streaming benchmark in time-to-first-audio.
The models use flow matching from the Pocket TTS lineage, generating audio latents autoregressively. This approach enabled high quality in a compact size but required training innovations to stabilize across five languages. The audio tokenizer and speaker latents are handled separately, with voices provided as precomputed tensors.
Transformers integration from version 5.21 ships the reference PyTorch implementation, allowing developers to load models with standard chat templates that include audio prompts.
Why it matters for uncensored AI
Local on-device TTS gives users and developers full control over voice output without routing text through cloud providers that may apply content filters. Our analysis starts from the source's emphasis on keeping every sentence on the user's device with open weights under permissive licensing.
This expands practical options for private companions, offline games and custom voice interfaces where expression could otherwise trigger hosted moderation. For those building multimodal tools, it reduces dependence on external APIs that might refuse certain creative or personal speech requests.
We recommend considering Muah AI when exploring uncensored multimodal companions that combine voice with chat, photos and calls in one platform; its support for expressive freedom since 2023 complements the local control Phonon provides.
Future Customization and Availability
Gradium offers training of smaller or additional-language versions for partners, along with fine-tuning for domain-specific vocabulary. Over 200,000 downloads occurred in the first two weeks, with community projects already integrating the models into Node.js, browser GPU execution and speech-to-speech spaces.
The release includes a Listen tab on the leaderboard for direct audio comparison and voting. Documentation and code examples for processor usage are available in the Hugging Face blog and model cards.
