How to Run Phonon-40M-en Locally: Gradium's 40M-Parameter Text-to-Speech Model

Phonon-40M-en is a 40 million parameter text-to-speech model from Gradium. It runs on CPU with a 74.6 MB quantized weight file, outputs 24 kHz mono audio, supports around 100 voices, and streams audio as it generates. License is CC BY-NC 4.0, so it is free for personal and research use.

Written by Priya Raghunathan, local-AI hardware reviewer. I verify parameter counts, benchmark numbers, and install commands against primary sources before publishing a setup path.

A new top-20 entrant on Hugging Face's text-to-speech trending list

Phonon-40M-en entered the top 20 of Hugging Face's text-to-speech trending list in October 2026. At time of writing it has 50 likes and 562 downloads on the gradium/phonon-40m-en model page, released October 19, 2026.

Specs at a glance

What it actually does

Phonon is a flow-matching transformer that turns text into speech. The model card says 40M parameters, which counts the transformer weights but not the Mimi codec decoder used to decode audio tokens into a waveform. The default quantized file is model.q8.gguf at 74.6 MB, which is small enough to ship inside a mobile app.

Text normalization runs automatically on numbers, currency, dates, and email addresses before synthesis. The model streams audio as generation proceeds rather than waiting for the full clip, which matters for real-time applications. It supports voice cloning through the Gradium studio if you need a custom speaker.

Benchmark performance

Phonon-40M-en scores 2117 on the Elo human-preference leaderboard, which measures listener preference in head-to-head comparisons. On intelligibility, it hits 2.3% word error rate on the hard set and 0.45% on the Seed-TTS set, both measured with an LLM-based evaluator. Lower WER is better.

Install it

Phonon-40M-en uses the ptts library, which pulls the quantized weights directly from Hugging Face on first run. Python 3.8 or newer works.

Generate speech from the command line

The ptts CLI downloads model.q8.gguf on the first call and caches it locally. Swap the --voice flag to any of the ~100 built-in speaker names.

Generate speech from Python

import ptts

tts = ptts.TTS("hf://gradium/phonon-40m-en", lang="en", quant="q8")
tts.save("speech.wav", "Hello from Phonon.", voice="Easton")

The TTS object streams audio internally. If you want to play audio while it generates rather than saving to disk, the ptts docs show a streaming callback API.

License

FAQ

What is Phonon-40M-en?

Phonon-40M-en is a 40 million parameter text-to-speech model from Gradium. It outputs 24 kHz mono audio, supports around 100 voices, and runs on CPU from a 74.6 MB quantized weight file.

Do I need a GPU to run it?

No. The model is built for on-device inference and runs on CPU. It is also designed to run in the browser and on phones. A GPU will reduce latency but is not required.

Can I use it commercially?

Not under the default license. CC BY-NC 4.0 covers non-commercial use only. For commercial licensing, contact Gradium directly.

Does it support languages other than English?

This checkpoint (phonon-40m-en) is English only. The en in the model name marks it as the English variant. Gradium also publishes separate French, German, Spanish, and Portuguese checkpoints, plus a multilingual variant.

How does text normalization work?

The model handles numbers, currency, dates, and email addresses automatically before synthesis. You do not need to expand "$12.50" or "Oct 19" in your input text.

Where to go from here

• Compare speech models side by side: the local AI model directory lists specs, licenses, and hardware requirements.