Ollama brings local speech to Macs, with different licenses by model

Ollama has introduced Phonon speech models for Apple Silicon Macs, allowing written text to be read aloud locally while audio is still being generated. Its October 24 announcement directs users to Ollama 0.40.11 or later. Windows and Linux speech-model support is planned, but the company has not given a release date.

Phonon receives text intended to be spoken. A chat model can feed its answer into the speech model, with playback starting after the first sentence or two rather than waiting for the entire response. The speech guide says terminal playback uses the computer where the command runs; redirecting its output saves a WAV file. The guide also describes file output on servers without speakers. This is text-to-speech, rather than a system that understands incoming speech.

Ollama reports 35% higher throughput than Gradium’s ptts runtime on an M5 Max MacBook Pro. It compared the median of ten runs generating 45 seconds of speech. Precision settings differed: Ollama used bf16, while ptts combined 8-bit CPU weights, fp16 on the Neural Engine and f32 on the GPU.

For accuracy, the company compared word error rates using 200 Seed-TTS English sentences transcribed by Whisper large-v3. It says its implementation matched the Hugging Face Transformers reference.

The default English model has 31 million parameters, excluding the audio tokenizer. Ollama lists loaded memory on the M5 Max as 121 MB, versus 141 MB for the 40-million-parameter variants. Those larger models cover English, French, German, Spanish and Portuguese, with separate language checkpoints and a multilingual option. Korean is not among the listed languages. Gradium lists the quantized English weights at 64.5 MB, a file size rather than loaded memory.

Licensing changes with the checkpoint. The 31-million-parameter English model is Apache 2.0, while the 40-million English, other language and multilingual checkpoints use CC BY-NC 4.0, limiting them to noncommercial use.

Gradium also publishes a separate catalog of 102 voice embeddings: 90 regular voices across seven locales and 12 English character voices. It describes ready-made embeddings used directly by Phonon, without a separate cloning step. Its Apache 2.0 license applies to those voice files; it does not replace the larger models’ noncommercial license.

Applications can receive generated audio through the OpenAI-compatible /v1/audio/speech endpoint. Its reference specifies 24 kHz, 16-bit mono WAV by default, or headerless raw samples for PCM. The speed field currently accepts only 1, and unsupported values return a 400 error. Input is limited to 4,096 characters per API request; the terminal command splits longer text at paragraph and sentence boundaries.