Gradium open-sources Phonon on-device TTS: a 31M-parameter, 64 MB model that starts speaking in 7 ms on a laptop CPU, with the English version under Apache 2.0

Background and the problem

Voice agents are spreading, but TTS still mostly runs through cloud APIs: every utterance needs a network round trip, metered billing and sending text to a server. On-device options have meant few languages, models too big for phone memory, or quality too low to ship. Kyutai's Pocket TTS made small on-device speech viable this year; Gradium announced Phonon in April, ran a private beta, and has now opened the weights.

Architecture and how it works

Phonon builds on Pocket TTS and Kyutai's Mimi codec: a flow-matching transformer predicts audio latents one 80 ms frame at a time and the Mimi decoder turns them into 24 kHz audio, streaming as it generates. The 31m count excludes the audio tokenizer. Built-in text normalization reads phone numbers, dates, temperatures and version strings correctly. Gradium ships its own runtime and q8 GGUF weights with CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU) and NPU (Core ML, Snapdragon) backends, plus Arm KleidiAI acceleration.

Benchmarks

All vendor-reported, with no independent replication yet: on a MacBook Air M5 CPU with 4 threads, Phonon 31m reaches first audio in 6.9 ms at 29.9x real time, versus 517 ms and 10.9x for Kokoro and 239 ms and 2.32x for NVIDIA Magpie (Kokoro does not stream). On a Raspberry Pi 5, 31m starts in 38 ms at 3.7x real time; on Snapdragon 8 Elite Gen 5, 7.5 ms at 24x. On Gradium's public 500-sentence hard set (gradium/tts-hard-sentences-202610), English WER is 1.70% for 31m, 2.11% for 40m multilingual, 2.22% for 40m, 3.50% for Kokoro, 7.40% for Kitten TTS and 8.71% for NeuTTS Nano. Naturalness was judged in blind pairwise tests with 1,189 raters and 17,270 votes against Kokoro, Magpie, NeuTTS and Pocket TTS.

Getting started

Try uvx ptts --lang en --quant q8 --voice en_us_Easton "Hello from Phonon." -o speech.wav, or pip install ptts and load ptts.TTS("hf://gradium/phonon-31m-en", lang="en", quant="q8") with streaming support. Browsers can use the phonon-tts npm package, and a Hugging Face Space lets you listen online. Gradium also points to a CPU-only Pipecat voice agent on Modal. For commercial products pick the English 31m (Apache 2.0); the 40m and multilingual models are CC BY-NC, non-commercial only unless licensed from Gradium.