Local speech generation starts with your machine and the model variant you choose. Check both before treating a working audio demo as a production option.
November 5: Phonon beyond the Ollama endpoint
Gradium’s November 5 technical follow-up describes Phonon’s ptts inference framework and its integration into Transformers from version 5.21. The framework targets several CPU, GPU and NPU backends. These runtime options are separate from the platform support of Ollama’s speech endpoint.
The follow-up distinguishes the English Phonon 31m model under Apache 2.0 from the five-language 40m option under a noncommercial license. A working runtime on a device does not establish commercial permission for every model variant. Keep the exact model card and terms with the chosen checkpoint.
For an integration plan, separate the model variant, inference runtime and consuming application. A browser or GPU backend described for ptts does not prove that the Ollama endpoint supports that same platform. This update adds the technical record; it does not report a new speech-quality or latency test.
Ollama adds a speech endpoint
Ollama’s October 24 announcement adds text-to-speech using Gradium’s Phonon models. The launch supports Apple Silicon Macs running Ollama 0.40.11 or later. Windows and Linux support is described as forthcoming.
The interface includes an OpenAI-compatible /v1/audio/speech endpoint. WAV is the default output, with 24 kHz mono audio. A raw 16-bit PCM option is also documented.
Endpoint compatibility can simplify integration with existing systems. You still need to check which request fields and output formats that integration expects.
Read the license for the exact variant
The Phonon model listing distinguishes the default English model from its smaller 40M variants. The default model uses Apache 2.0. The 40M language variants use CC BY-NC 4.0, which includes a noncommercial restriction.
The listing includes French, German, Spanish, Portuguese and multilingual 40M options. A language match does not establish permission for a commercial workflow. Review the actual variant’s terms before selecting it.
Keep the model tag and license record with the application configuration. A later change to a smaller or multilingual variant can change the terms that apply.
Test the audio where it will be used
I would use a short passage containing the names, numbers, and abbreviations that readers actually encounter. Listen to the whole output on the intended device.
- Confirm the Mac architecture and installed Ollama version.
- Record the precise model variant and applicable license.
- Check names, dates, punctuation, and pronunciation.
- Measure startup delay separately from sustained generation.
- Verify that the consuming player accepts the returned format.
For a read-aloud feature, include a correction path for mispronounced names. A clean waveform or successful HTTP response can’t prove the spoken information is accurate.
Keep the speed claim in its benchmark context
Ollama reports a throughput improvement in its own comparison. The test uses a particular Mac, speech sample, and runtime configuration. Different precision settings and hardware can change the result.
We have not measured the models’ latency or voice quality. The practical comparison is the time needed to produce acceptable speech on the machine and model variant that your workflow can use.
The VoiceStudio guide covers another local audio rights check. Our Voxtral 3 article addresses interruptions and tool actions in a full voice-agent workflow.
Sources
Does the exact model variant fit both your language requirements and your intended use?
Prepared with AI-assisted research and drafting. MustHave.ai has not performed the proposed product tests.



