Gradium open-sources Phonon on-device TTS: English 31M under Apache 2.0, 64.5 MB offline
MarkTechPost reported on October 22, 2026 that Gradium open-sourced Phonon, an on-device text-to-speech family. The English Phonon 31m has 31M parameters, 64.5 MB q8 weights, and an Apache 2.0 license. Gradium's internal tests put its WER at 1.70% on a difficult sentence set and time to first audio at 6.9 ms on a MacBook Air M5 CPU.
What changed
According to a MarkTechPost report dated October 22, 2026, Paris-based speech company Gradium, spun out of Kyutai, open-sourced the on-device text-to-speech model family Phonon. The family contains three models: English-only Phonon 31m, with 31M parameters (excluding the audio tokenizer), 64.5 MB q8 GGUF weights, 24 kHz mono output, 37 built-in voices, and an Apache 2.0 license; separate per-language Phonon 40m models (English, French, German, Spanish, Portuguese) at 74.6 MB; and a Phonon 40m multilingual version that packs the five languages into a single model with 102 voices, the latter two under the non-commercial CC BY-NC 4.0 license. Architecturally, Phonon uses a causal transformer backbone with text and speaker embeddings concatenated as a prefix, a 4000-token SentencePiece tokenizer that reads raw text directly without phoneme conversion, and a backbone that outputs one latent vector per 80 ms frame. A flow-matching head converts that into an audio embedding in a single step, a scaled-down variant of the Kyutai Mimi decoder produces 24 kHz PCM, and a separate end-of-speech head decides when to stop. Phonon 31m has 6 layers, width 512, and 8 attention heads; the 40m has 9 layers at the same width. Both are distilled from a larger Phonon teacher model and were trained on NVIDIA H100 clusters, building on Kyutai's continuous audio language model research and Pocket TTS. Performance figures released by Gradium include: in a blind test with 1,189 raters and 17,270 votes, American English Phonon 40m scored 2117 Elo and Phonon 31m 2098, versus 2043 for Kokoro; on a public difficult sentence set (500 sentences per language, including phone numbers, email addresses, and postcodes), Phonon 31m had a WER of 1.70%, compared with 3.50% for Kokoro and 8.71% for NeuTTS Nano; on the easier Seed-TTS set the gap narrowed, with Phonon 31m at 0.51% LLM WER, slightly worse than Kokoro at 0.45%. On speed, with 4 threads on a MacBook Air M5 CPU, time to first audio was 6.9 ms at about 29.9x real time (up to 46.7x with Arm KleidiAI and Apple Accelerate); on a Raspberry Pi 5, 38 ms startup at 3.7x real time; on a Snapdragon 8 Elite Gen 5, 7.5 ms at 24x real time; on an NVIDIA L40S, 5 ms with CUDA; and on an AMD EPYC 7413 CPU, 43 ms. The runtimes are MIT or Apache-2.0 licensed and cover Rust, Python (pip install ptts), Swift, C, C++, and the browser (phonon-tts), with an OpenAI-compatible server and a WebSocket server also provided. The models can be loaded in Transformers, MLX-Audio, Ollama, and LM Studio; voice cloning and voice design are done in Gradium Studio and then run locally.
What it means for developers
For developers, the practical value of this news is turning English TTS from a per-character cloud call into a local component that can ship with an application. At 31M parameters and 64.5 MB of q8 weights, it can be bundled with business code into a mobile install package or browser assets, without depending on a network round trip and without consuming a GPU budget. Streaming brings time to first audio in the 6.9 ms range, making latency-sensitive use cases such as real-time dialogue and interactive read-aloud feasible on CPU alone. Three caveats matter, though. First, Apache 2.0 covers only English Phonon 31m; the 40m and 40m multilingual versions are CC BY-NC 4.0, so a commercial product cannot treat multilingual capability as a free extension under the same license. If French, German, or other languages are needed, either seek separate licensing or keep using another option. Second, the Elo, WER, and speed figures in the report all come from Gradium's own evaluations, with the benchmark sets and comparison targets chosen by the vendor. The lead of 1.70% WER on the difficult sentence set over Kokoro's 3.50% is attributed mainly to built-in text normalization, while on the more conventional Seed-TTS set Phonon 31m was slightly behind Kokoro, indicating the advantage is concentrated on structured text such as phone numbers, email addresses, and postcodes and should not be extrapolated into overall superiority. Third, quantized size, multi-thread latency, and throughput depend heavily on the specific hardware, thread count, and runtime backend; the same model can behave noticeably differently across CPUs and binding methods, so the single-point numbers in the report are only useful as an order-of-magnitude reference.
Integration and evaluation
Two conditional paths are worth pursuing. First, if the product's main need is offline English narration and the deployment target is CPU or mobile, consider Phonon 31m directly as a candidate replacement for cloud TTS: first confirm the license scope applies only to English 31m, then test q8 weights on the target device. A reasonable evaluation criterion is to run a batch of real business text on the target device with a fixed thread count, measure P50 and P95 time to first audio, and compare them with the end-to-end latency (including network) of the existing cloud option; if P95 time to first audio does not meet an acceptable threshold, fall back to the cloud option. Second, if the product needs multiple languages, treat the 40m multilingual version as a separate non-commercial option: use it only for internal prototypes until the business model is confirmed to permit CC BY-NC. A reasonable criterion is to compare per-language WER and subjective voice naturalness scores using multilingual samples from the difficult sentence dataset, while recording whether the license satisfies distribution requirements; if either fails, keep the existing vendor. Third, whichever path is chosen, run a text normalization comparison first: extract structured text such as phone numbers, email addresses, postcodes, and amounts into a fixed test set, run the same batch through Phonon with its built-in normalization and through options that require custom preprocessing, and compare error rates, because the report explicitly identifies this as the main source of the gap; if in-house normalization is already good enough, Phonon's advantage here disappears and selection should shift to latency and size. Fourth, include the runtime backend as a variable in evaluation: compare the plain CPU path with versions enabling acceleration libraries such as Arm KleidiAI and Apple Accelerate to confirm whether the order-of-magnitude throughput gain can be reproduced on the target chip; if not, plan concurrency capacity using conservative values and assess in advance the operational cost of the WebSocket and OpenAI-compatible servers in self-hosted scenarios. All of the above are recommendations and conditional analysis based on the report, not verified facts.
Limits and open questions
The Elo, WER, and latency figures in the report all come from Gradium's own evaluations and press materials; the benchmark sets, comparison targets, and test conditions were chosen by the vendor and have not been independently reproduced by a third party. The fact that Phonon 31m was behind Kokoro on the Seed-TTS set also shows the advantage is conditional. On licensing, only English Phonon 31m is confirmed as Apache 2.0, while the 40m and multilingual versions are CC BY-NC 4.0; open weights do not equal free commercial use, and before selection you should check the original model cards rather than infer rights from the mere availability of downloadable weights. In addition, Phonon 31m and the entire Phonon family do not currently appear in KimiSeek's live model catalog. The availability, licensing, and performance information cited here comes from Gradium and the report; it is unrelated to the KimiSeek platform and does not mean the platform offers this model or a related API. Prices and license terms from various parties may change over time, so refer to the latest pages from the publisher and the platform.