What happened
In a November 5, 2026 post, Gradium describes the release of Phonon, a family of open-weight text-to-speech models. The company says it announced Phonon in April and spent the following six months developing it with partners and early users. Two model sizes are available, about 65 MB and 75 MB. Together, the models support English, French, German, Spanish and Portuguese, with more than 100 built-in voices. Users can also create voices through Gradium Studio, including by cloning a voice from a short recording with the speaker’s consent or designing one with a text description.
The release’s main deployment claim is that Phonon can run on phones, laptops, single-board computers and in a browser tab, without a server or network connection. Gradium says this keeps the text on the user’s device. The models are based on the Pocket TTS lineage and use flow matching to generate audio. The company says adapting the approach to small models required addressing numerical instability during training. Phonon also handles items such as phone numbers, dates and codes, according to the post.
Licensing depends on the version. Phonon 31m, which speaks English, is under Apache 2.0 and can be used in commercial products. Phonon 40m covers all five languages under a non-commercial CC BY-NC license. Gradium says commercial projects using the latter can contact the company for pricing based on user numbers rather than speech volume. The post does not specify a public price. Gradium also offers to train custom Phonon models with design partners, including versions for other languages, tighter memory budgets or product-specific vocabularies.
For inference, Gradium built ptts, a framework intended to run Phonon across more than ten backends. Its listed targets include Intel and Arm CPUs, WebAssembly, CUDA, Vulkan, Metal, WebGPU, Apple Core ML and Snapdragon. The post says ptts is oriented toward running a single stream on-device. Phonon is also available in Hugging Face Transformers from version 5.21, where the reference PyTorch implementation targets GPU throughput through batched generation and torch.compile. The source describes the model as autoregressively generating audio latents, which a processor decodes into audio. It reports 30,788,354 parameters for the Phonon 31m model and 10,305,313 for its audio tokenizer.
Gradium reports more than 200,000 downloads in the first two weeks. On Hugging Face’s Open TTS Leaderboard, the company says Phonon ranks first in intelligibility for German and Spanish and in the top three for every language it evaluated, alongside models up to 100 times larger. It also says Phonon tops the CPU streaming benchmark on time-to-first-audio. The leaderboard evaluates intelligibility, speed and naturalness, with separate treatment of fixed voices and voice cloning. The post points to early projects on devices, in browsers and in Node.js, as well as a Transformers integration built by Hugging Face.
What the archive adds
Gradium open-sourced Phonon, a family of 31M-40M parameter on-device TTS models covered the earlier open release and the shift toward embedding speech generation in applications. This is a new fact on a known story: the current post adds the five-language release details, licensing split, runtime support, reported benchmark positions and first-two-weeks download figure.
Trend signal
This release supports the pattern that generative media is becoming application infrastructure: Generative media becomes application infrastructure. Phonon is presented not only as a model to query, but as a component that can run inside applications across device classes and software runtimes. That is evidence of the integration direction, not proof that local TTS is suitable for every production workload.
Why now
Gradium says it spent six months iterating with partners and early users after announcing Phonon in April. The November post explains what changed in that development period and shares the model’s early download and leaderboard results. It also arrives with Transformers support and several community and team demos, giving developers a clearer view of how the release can be tried and integrated.
Connections to other events
The archive’s earlier Telegram coverage of AssemblyAI’s Voice Standard Bench described comparisons of models on tasks performed after speech recognition, while its October 27 item on Cekura covered latency, pronunciation, naturalness, cost and interruption behavior for voice-agent TTS. Phonon meets that broader evaluation and deployment thread from a different angle: it emphasizes local inference and compact model size rather than a general ranking of hosted voice services. The source itself also situates Phonon in the Pocket TTS lineage.
Why it matters
For teams building speech into applications, an offline model changes the deployment question. Text does not need to make a round trip to a provider, and speech generation can remain available where connectivity is absent. That can matter for privacy-sensitive experiences, embedded devices, browser applications and products that need predictable local availability. The reported sizes and broad backend list make experimentation plausible on hardware beyond a server GPU, although the source does not provide per-device latency, memory use or battery measurements.
The release also makes licensing a product decision rather than a footnote. The English-only 31m model’s Apache 2.0 license is materially different from the five-language 40m model’s non-commercial license. A team evaluating multilingual commercial deployment therefore cannot assume that the most convenient checkpoint is cleared for its use. Builders should validate actual voice quality and performance, and confirm permissions, on the exact model they intend to ship.
What happens next
- More community integrations will appear for Phonon across browser and edge runtimes. (6 months, 65%). Watch for: additional public projects or demos named in the model’s Hugging Face materials.
Limits and uncertainty
These results come from Gradium’s launch post, which presents its own model and reports its leaderboard standing. The post gives no full per-language scores, hardware configuration, latency values, memory footprint or power measurements, so the top placements do not establish how Phonon will perform on a particular device or against a particular commercial system. The download count is reported by the company and does not show active use or production adoption. The source also does not publish commercial pricing for the CC BY-NC models, nor does it establish that every custom voice workflow or deployment is appropriate without consent and licensing review.
Skill to build now
Take a small set of representative product phrases—including dates, phone numbers, codes and names—and test them in Phonon on the device you plan to support. Compare intelligibility, startup time and memory use with your current speech service while offline. If the use is commercial, first confirm that the selected checkpoint’s license covers the product and language combination.
Reader takeaway
Phonon makes local, multilingual text-to-speech more concrete: it is compact, has multiple runtime paths and has early benchmark results. But the practical decision is not simply whether it ranks well. Test it on your target hardware and verify the specific model’s license, especially if you need the five-language version commercially.