Phonon fits local TTS into 65 MB, but its licenses split the model family

Phonon packages local TTS in two sizes of roughly 65 MB and 75 MB. The models cover five languages, run on phones, computers and in browsers, and can synthesize speech without sending text to a server.

Two small models cover five languages and more than 10 backends

Phonon 31m has 30,788,354 parameters excluding the audio tokenizer, supports English and ships as a 64.5 MB 8-bit file. Phonon 40m has variants for English, French, German, Spanish and Portuguese at about 74.6 MB each. The voice catalog contains more than 100 options.

The ptts runtime supports over 10 backends. They include CPU execution for Intel, Arm and WebAssembly, GPU paths through CUDA, Vulkan, Metal and WebGPU, plus NPU support through Apple Core ML and Snapdragon. A reference implementation is available in Transformers from version 5.21.

Local speech removes the network trip, while licensing controls deployment

On-device execution suits accessibility tools, games, offline assistants and products that cannot send text to a cloud service or wait for a server response. Streaming can also start playback before the model has completed the full input.

The important dividing line is licensing. The English Phonon 31m uses Apache 2.0 and allows commercial use. The 40m models use CC BY-NC 4.0 and are free for non-commercial projects. Five-language coverage therefore does not provide the same deployment freedom across the family.

Leaderboard results come from the authors and cannot cover every product

The team reports first place in German and Spanish and a top-three result in every evaluated language on the Hugging Face Open TTS Leaderboard. It also claims the best CPU streaming time to first audio. The leaderboard is new, however, and the vendor is reporting its own interpretation of the results.

The model card lists practical limits. Long inputs are split into sentences, so prosody does not carry across the full text. Language selection must match the text normalizer, and unsupported languages receive incorrect pronunciation. Each product still needs to test names, addresses, domain abbreviations and its target hardware.

Target-device tests and voice governance will decide real deployments

The first test is latency, memory and battery use on an actual phone or embedded device rather than a reference machine. The second is accurate handling of numbers, dates and codes in the product's language.

Phonon supports voice cloning from a short recording and asks users to obtain the speaker's consent. Products will need their own provenance rules, synthetic-speech disclosure and safeguards against impersonation. A smaller model lowers the technical barrier. It does not remove responsibility for the voice.

Lilith's verdict

Phonon has compressed a voice model until it can sit beside an ordinary mobile app. The language picker now needs a second button, though: the license decides whether that voice may also speak for a paying product.