Text to speech that runs on your own Mac: Gradium's Phonon, served by Ollama. Its first audio arrives in about 30 ms, and it speaks 35–45× faster than real time. This repo shows the model off five ways (from the terminal, a Markdown file read aloud, a voice gallery, a benchmark, and a local web page), in plain Python with no runtime dependencies.
An independent demo, not affiliated with Gradium or Ollama. Measured on 2026-10-25; the model is a week old, so re-check anything you rely on.
| what it does | text in, speech out: 16-bit mono WAV or raw PCM at 24 kHz, streamed while it is generated |
default tag phonon |
43M parameters, 112 MB, 38 English voices, Apache 2.0 |
phonon:40m-multilingual |
57M parameters, 148 MB, 103 voices across English, French, German, Spanish and Portuguese, CC BY-NC 4.0 (non-commercial) |
| runs on | Apple Silicon only, through Ollama's MLX runner; needs Ollama 0.40.9 or later |
| API | Ollama's OpenAI-compatible POST /v1/audio/speech, at most 4,096 characters per request |
The voices range from plain (Zoey, Easton, Freya, Hugh) to characters (Captain Silas Blackwater, Grub the Ogre, The Snow Queen, The Thing Under the Stairs). phonon-demo voices lists them. Eight are in samples/, all reading the same sentence.
brew upgrade ollama # Phonon needs 0.40.9+; `ollama --version` to check
ollama pull phonon # 112 MB
git clone https://github.com/gsaini/phonon-speech-demo.git
cd phonon-speech-demo
uv sync # or: python -m venv .venv && .venv/bin/pip install -e .
uv run phonon-demo say "Good morning. Checkout recovered at 14:32." --voice Zoey
uv run phonon-demo serve # the web page above, on 127.0.0.1:8766| command | what it does |
|---|---|
say "text" |
speaks it, playing as the audio streams in (through ffplay if installed, else afplay at the end), and prints the timings. --out file.wav or .mp3 saves it too; - reads stdin. |
read notes.md |
reads a file aloud: Markdown formatting, code blocks, URLs and badges dropped, headings and table rows turned into sentences, the text split at sentence breaks into requests of at most 4,096 characters and joined back into one clip. --show prints what would be spoken. |
voices |
the model's voices |
gallery |
one sample per voice into a folder, with an index; --voices Zoey Easton for a few, --format mp3 |
bench |
time to first audio and speed versus real time, cold and warm, for a sentence, a paragraph and a page |
serve |
the web page: type, pick a voice, hear it, see its waveform, replay earlier takes. /?text=…&voice=Zoey prefills it, so a phrase can be shared as a link. |
--model phonon:40m-multilingual switches any command to the multilingual checkpoint (ollama pull it first), and --host or $OLLAMA_HOST points at another Ollama.
uv run phonon-demo bench on an M1 Pro with 16 GB, Ollama 0.40.11, the default phonon tag. Warm rows are the median of three runs.
| text | characters | audio | first audio | done after | faster than real time |
|---|---|---|---|---|---|
| one sentence, cold | 58 | 3.8 s | 363 ms | 439 ms | 9× |
| one sentence | 58 | 3.7 s | 26 ms | 100 ms | 36× |
| one paragraph | 272 | 18.4 s | 26 ms | 419 ms | 44× |
| a page | 2,532 | 148.5 s | 28 ms | 3,379 ms | 44× |
- Time to first audio doesn't grow with the text. About 26–30 ms for a sentence or a page, because the audio streams as it is generated. That is what makes it usable for a voice that answers as an assistant writes.
- It is 35–45× faster than real time once loaded, with every voice tried: the eight samples were each made in 0.12–0.18 s. Reading the README of system-one-incident-triage aloud, about seven minutes of speech, took 9.5 s in two requests.
- Cold costs a few hundred milliseconds. The cold row is the first request after
ollama stop phonon. The very first load right afterollama pulltook about 4 s, presumably while the weights were read from disk for the first time.
The endpoint is OpenAI-compatible, so any client works:
curl http://localhost:11434/v1/audio/speech \
-d '{"model": "phonon", "input": "Hello there.", "voice": "Zoey"}' --output hello.wavfrom openai import OpenAI
client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
with client.audio.speech.with_streaming_response.create(model="phonon", voice="Zoey", input="Hello there.") as response:
response.stream_to_file("hello.wav")Or this repo's client, standard library only:
from pathlib import Path
from phonon_demo.audio import save
from phonon_demo.client import Phonon
speech = Phonon().speak("Any length of text: it is split and joined for you.", voice="Freya")
save(speech.pcm, Path("out.wav"))
print(f"{speech.seconds:.1f} s of audio, first audio after {speech.first_audio_ms:.0f} ms")- Ask for
pcm, write your own WAV header. Thewavstream's header is written before the length is known, so its size fields hold placeholder maximum values. Most players cope; some editors and parsers don't. The client here requests raw PCM and writes a correct header (audio.py). - 4,096 characters per request.
ollama run phononsplits long text itself; the HTTP endpoint doesn't, and returns 400. text.py splits at paragraphs, then sentences, then words. - Clean Markdown before speaking it. The endpoint skips Markdown markers and emoji, but code blocks, link targets and badges are still words. A README's badge row comes out as "CI, Python, Offline, License" unless images are dropped.
- An unknown voice is a 400 that lists the valid ones, so the error message is a usable voice list.
- Licenses differ by checkpoint. The default English model is Apache 2.0. The 40M checkpoints, including the multilingual one, are CC BY-NC 4.0: fine for a demo, not for a product. This repo includes no model weights. The samples in samples/ were made with the Apache-licensed default.
src/phonon_demo/
client.py Phonon: version and capability checks, voices, streaming PCM, long text split and joined, timings
text.py speakable(): Markdown to plain sentences; chunks(): pieces of at most 4,096 characters
audio.py WAV with a correct header, .mp3 through ffmpeg, streaming playback through ffplay
bench.py cold and warm timings over three lengths of text
cli.py `phonon-demo say | read | voices | gallery | bench | serve`
web/ a standard-library server on 127.0.0.1 and one page; the waveform is drawn from the returned audio
samples/ eight voices, one sentence each, as MP3
tests/ 20 offline tests against a fake Ollama, so CI needs neither Ollama nor Apple Silicon
MIT for the code. The Phonon models carry their own licenses, above.