Phonon
A series of on-device text-to-speech (TTS) models open-sourced by Gradium, emphasizing small parameter counts, low latency and commercial usability, and intended for deployment on local devices.
Timeline (last 90 days)
The Phonon voice catalog lets users select a specific voice by name or identifier, rather than having to guess which timbre corresponds to a model file themselves.
Ollama v0.40.11 added voice catalog support to the Phonon speech engine.
Phonon is the name of an engine/component internal to Ollama used to carry speech capabilities, loaded and scheduled by mlxrunner (a model runner based on the Apple MLX framework); it is not a publicly known third-party speech model brand.
The voice catalog was introduced and listed by mlxrunner, indicating that the current implementation focus of Phonon speech capabilities is the Apple platform.
For teams building local speech applications, the direct value of the voice catalog is predictability, reducing trial-and-error costs during debugging.
How many timbres the Phonon voice catalog can ultimately cover and which languages it supports are not elaborated in the commit itself and should be subject to the documentation and subsequent versions.
The change also updated the documentation, elevating the Phonon voice catalog from an internal implementation to part of the public commitment.
The minor version number of the v0.40.11 version line has reached two digits, indicating that the project is still iterating rapidly.
Tag v0.40.11-rc0 appeared on GitHub, with the commit message "voice: add voice catalog for Phonon (#19242)".
The commit contains two substantive changes: "mlxrunner: import and list Phonon catalog voices" and "docs: document Phonon voices".
5 more timeline events
All Facts (15)
The voice catalog was introduced and listed by mlxrunner, indicating that the current implementation focus of Phonon speech capabilities is the Apple platform.
50%UnverifiedOllama v0.40.11 added voice catalog support to the Phonon speech engine.
50%UnverifiedPhonon is the name of an engine/component internal to Ollama used to carry speech capabilities, loaded and scheduled by mlxrunner (a model runner based on the Apple MLX framework); it is not a publicly known third-party speech model brand.
50%UnverifiedThe change also updated the documentation, elevating the Phonon voice catalog from an internal implementation to part of the public commitment.
50%UnverifiedThe Phonon voice catalog lets users select a specific voice by name or identifier, rather than having to guess which timbre corresponds to a model file themselves.
50%UnverifiedFor teams building local speech applications, the direct value of the voice catalog is predictability, reducing trial-and-error costs during debugging.
50%UnverifiedHow many timbres the Phonon voice catalog can ultimately cover and which languages it supports are not elaborated in the commit itself and should be subject to the documentation and subsequent versions.
50%UnverifiedThe minor version number of the v0.40.11 version line has reached two digits, indicating that the project is still iterating rapidly.
50%UnverifiedThe commit contains two substantive changes: "mlxrunner: import and list Phonon catalog voices" and "docs: document Phonon voices".
50%UnverifiedThe import and listing functionality for the Phonon voice catalog is implemented in the mlxrunner module.
50%UnverifiedThe scale of 31M parameters and 64 MB means the vocoder and embedding tables have been significantly compressed, possibly using parameter quantization (such as int8), knowledge distillation or a lightweight vocoder approach.
50%UnverifiedThe original article did not disclose Phonon's audio quality evaluation data, the number of supported languages, inference framework dependencies or hardware compatibility scope.
50%UnverifiedThe 31M parameter scale is roughly only one percent of mainstream cloud TTS models.
50%UnverifiedTag v0.40.11-rc0 appeared on GitHub, with the commit message "voice: add voice catalog for Phonon (#19242)".
50%UnverifiedGradium has open-sourced the Phonon series of on-device text-to-speech (TTS) models.
50%