Phonon TTS, two weeks on: top CPU streaming latency, but Apache 2.0 covers only the English model
According to a Gradium team post on the Hugging Face blog dated November 5, 2026, the on-device TTS model Phonon followed up on its release details: a 31m English version under Apache 2.0 and a 40m five-language version under CC BY-NC, self-reported first place in German and Spanish intelligibility and in CPU streaming time-to-first-audio, plus Transformers 5.21 integration and a ptts inference path.

What changed
Published November 5, 2026, the post is Gradium's retrospective on Phonon, the on-device speech model whose weights it opened two weeks earlier. The team says Phonon comes in two sizes, roughly 65 MB and 75 MB, which together cover English, French, German, Spanish and Portuguese, ship with more than 100 built-in voices, allow custom voices to be created in Gradium Studio, and handle segments commonly misread aloud such as phone numbers, dates and code. The license is split in two: the English Phonon 31m uses Apache 2.0 and can be used in commercial products; the five-language Phonon 40m uses CC BY-NC, free for non-commercial projects, with commercial licensing priced by user count. The model continues the Pocket TTS line, generating audio with flow matching; the team says training was numerically unstable early on and needed several techniques before reaching high quality across five languages. The post also lists results on the Hugging Face open TTS leaderboard: first place in German and Spanish intelligibility, top three in every language evaluated, and first place in time-to-first-audio on the CPU streaming benchmark, with figures given for the 95th percentile at 28 ms for the 31m English version, 38 ms for the 40m English version and 165 ms for Pocket TTS. On integration, Phonon has been added as a reference PyTorch implementation starting with Transformers 5.21, and there is a separate ptts inference framework designed for multi-device backends, supporting more than ten backends across CPU, GPU and NPU; voices are not recordings but precomputed speaker latent tensors shipped in the gradium/phonon-voices repository, and the audio tokenizer is released in decoder-only form. The team says downloads exceeded 200,000 in the first two weeks and mentions community projects such as in-browser ExecuTorch conversion and a Node.js-side Micdrop integration.
What it means for developers
The following is analysis based on the report above, not verified fact. The impact on developers centers on three points. First, the license splits by version, and picking the wrong version is harder to undo than picking the wrong model: for English-only commercial use, 31m is the one with a clear license; if you need German, French, Spanish or Portuguese output, the 40m's CC BY-NC means the weights can be downloaded but commercial use is restricted, with commercial licensing priced by user volume, something to negotiate in advance rather than patch in before launch. Second, the same weights appear in two inference forms: per-sentence interaction on device goes through ptts single-stream, while offline batch synthesis goes through Transformers batched generation with full-graph compilation; these are a division of scenarios, not substitutes, and choosing wrong costs you clearly on either latency or throughput. Third, voices are delivered as precomputed tensors, and the released tokenizer can only decode, not directly encode your own audio, so custom voices must go through the Gradium Studio process with speaker consent; managing voice assets differs from the old "upload a recording" path. It is worth stressing that leaderboard rankings and millisecond figures come from the vendor's post and measurements on the leaderboard under specific streaming CPU conditions, and may not reproduce on a given phone, browser or single-board computer.
Integration and evaluation
The following are conditional suggestions, not verified conclusions. Suggestion one: settle versions by language and license first. A testable criterion is a recognition round-trip on a fixed sentence set from your business (including phone numbers, dates and code), measuring word error rate for 31m and 40m in the target languages, and checking the license file of the version you will distribute; if you are English-only and commercial, testing 31m is enough to proceed, but if the target languages include German, French, Spanish or Portuguese you must settle the commercial licensing path before investing in integration, because the 40m's non-commercial license will block release directly, and this condition determines the whole order of selection. Suggestion two: choose the inference path by latency or throughput. If the product reads sentences aloud on a device and users wait for the first audio, choose ptts and set the 95th-percentile time-to-first-audio on your target hardware as the acceptance metric, treating the post's leaderboard numbers only as a reference baseline; if you need offline batch generation of speech datasets, use the Transformers path, switch the acceptance metric to samples per second and memory use on fixed hardware, and re-measure with batching and full-graph compilation enabled, because the two paths optimize for opposite goals. Suggestion three: treat voices as assets that must be generated externally. If you need custom voices, first confirm that Gradium Studio's cloning and text-design flow works and that the speaker consent chain can be closed; as an evaluation metric, you can follow the leaderboard's approach and score speaker similarity between cloned and built-in voices on your own evaluation set, falling back to built-in voices if similarity falls short or consent cannot be closed; whether this precondition holds determines whether you can promise a distinctive voice externally.
Limits and open questions
KimiSeek's live model catalog currently contains neither Phonon 31m nor Phonon 40m; availability mentioned here refers to the vendor side, meaning weights published on Hugging Face and Gradium Studio, not a KimiSeek service, and there is no evidence that either version is offered to developers as an API. Leaderboard results and the 28 and 38 ms time-to-first-audio figures come from Gradium's own post and the Hugging Face leaderboard under specific streaming CPU conditions; they are vendor claims and were not independently reproduced here. The post gives no per-language latency distributions, memory-use curves or long-run stability data; the specific terms of the 40m's CC BY-NC restriction and commercial licensing need confirmation from the vendor; quantized smaller-memory versions, Korean, Arabic and Japanese support, and fine-tuning on product vocabulary are all described only as under discussion, with no timeline. Open weights are not the same as an open-source license: the 40m is exactly the case of weights being available while the license is non-commercial, and only the 31m's Apache 2.0 is an open license permitting commercial use.
Illustrations from the original report
