Gradium Phonon 31M Ships 6.9 ms CPU TTS Under Apache 2.0

Gradium released Phonon on October 19, 2026, open-sourcing a family of compact text-to-speech models aimed at CPU-class hardware. The 31M-parameter English variant, excluding the audio tokenizer, ships under Apache 2.0, with 64.5 MB q8 GGUF weights and a 6.9 ms first-audio time Gradium reports on an M5 CPU. Those two numbers remove the most common blockers for local speech: edge storage budgets and cloud round-trip latency. As covered in choosing local AI workloads, on-device inference decisions are increasingly driven by data boundaries and latency rather than raw capability gaps.

Architecture: flow-matching on 80 ms frames

Phonon avoids the two-stage phoneme pipeline common in older TTS systems. A SentencePiece tokenizer with a 4,000-token vocabulary reads raw text directly; Gradium reports this simplifies adding languages. Tokenized text and a speaker embedding feed a causal transformer — 6 layers, width 512, and 8 attention heads in the 31M variant — that generates one latent vector per 80 ms audio frame autoregressively. A flow-matching head converts each latent plus noise into a continuous audio embedding in one step, and a smaller variant of Kyutai's Mimi codec decoder turns the embeddings into 24 kHz mono PCM. A separate end-of-speech head decides when to stop. The 40M variant uses 9 layers at the same width. Both are distilled from a larger Phonon teacher and build on Kyutai's Continuous Audio Language Models research and the earlier Pocket TTS system. Training used an NVIDIA H100 cluster with NCCL and CUDA Graphs.

Benchmarks: structured-text accuracy and listening preference

All benchmark figures are vendor-reported by Gradium; the English Elo comes from 468 raters and 8,290 blind pairwise votes.

The hard-set WER gap is largely driven by built-in text normalization for phone numbers, postal codes, dates, and units. On the less structured Seed-TTS set, Gradium reports phonon-31m-en trails Kokoro at 0.45% and Magpie at 0.46%, as MarkTechPost also noted. MarkTechPost notes the model card lists 1.5% while the Gradium blog lists 1.70% for the same 500-sentence evaluation.

Latency and hardware targets

Gradium reports phonon-31m-en reaches first audio in 6.9 ms on a MacBook Air M5 CPU with 4 threads and no KleidiAI optimization, running 29.9× faster than real time. With Arm KleidiAI and Apple Accelerate, throughput rises to 46.7× and startup drops to 5.4 ms. A Raspberry Pi 5 starts in 38 ms at 3.7× real time; a Snapdragon 8 Elite Gen 5 Oryon CPU starts in 7.5 ms at 24×, measured by Gradium with Qualcomm. Gradium reports an NVIDIA L40S reaches 5 ms via the CUDA backend. The runtime spans Rust, Python (pip install ptts), Swift, C/C++, and the browser via WASM/WebGPU, and it integrates with Transformers, MLX-Audio, Ollama, and LM Studio. An OpenAI-compatible Docker server is also provided.

AI Mastery analysis

The license split inside the Phonon family is a sharper deployment constraint than the headline Elo numbers suggest. Phonon-31m-en under Apache 2.0 is the only variant cleared for unrestricted commercial products; the 40M English model and every multilingual model are CC BY-NC 4.0. Teams needing French, German, Spanish, or Portuguese synthesis — or the better Seed-TTS WER of the 40M English model — must accept the non-commercial restriction or negotiate separately with Gradium. The 31M model also documents a prosody-reset limitation on long text, which matters for narration and document-to-speech pipelines.

The 6.9 ms startup figure is hardware-specific. It applies to an M5 CPU, not to general x86 servers; Gradium reports 43 ms on an AMD EPYC 7413, a different backend target. Even so, Gradium's comparison puts Kokoro at 517 ms on the same M5 class, a gap that follows from Phonon streaming frame-by-frame rather than decoding a complete utterance before playback begins. This is the same trade-off covered in constrained inference: committing to smaller output units early changes the latency profile. The 80 ms frame is the atomic unit, and it is why the model can plausibly run on a Raspberry Pi 5 without buffering.

The permissive-licensed 31M release, sub-100 MB footprint, and runtimes from C headers to Android TTS integration lower the floor for production-quality on-device speech. The remaining ceilings are conversational prosody and the commercial license gap on multilingual variants.

Sources

• Primary: gradium/phonon-31m-en · Hugging Face
• Reported by: MarkTechPost: Gradium Open-Sources Phonon, a 31M-Parameter On-Device TTS Model Under Apache 2.0