Phonon Brings Fast On-Device Speech to More Than 200,000 Downloads

Gradium says its Phonon on-device text-to-speech models passed more than 200,000 downloads during their first two weeks. The company shared the figure on November 5, 2026, in a technical follow-up published on Hugging Face.

The open-weight release arrived on October 20, 2026, giving developers access to speech models designed to run on the user’s device instead of sending text to a cloud service. Phonon produces mono 24 kHz audio, can stream speech while generation continues and offers more than 100 built-in voices.

That combination targets offline assistants, accessibility tools, games and other interactive applications that cannot wait for a cloud round trip. It also keeps text on the user’s device, which makes the deployment model part of the product’s appeal.

Strong leaderboard results across multiple languages

Gradium said Phonon ranked first for intelligibility in German and Spanish on Hugging Face’s Open TTS Leaderboard. The company also said Phonon placed in the top three in every language it evaluated, alongside models up to 100 times its size.

The evaluation covered 23 languages and looked at intelligibility, inverse real-time factor, time to first audio and automated naturalness measures. Voice-cloning tests added speaker similarity, giving the evaluation a separate way to examine whether generated speech preserved the identity of a reference voice.

Hugging Face calculates intelligibility by comparing automatic speech recognition transcripts with the original prompts. The comparison uses word or character error rates. Streaming tests run one utterance at a time with common prompts and the same hardware, while models without a streaming interface must generate the entire utterance before playback can begin.

That method offers a useful comparison, but it has limits. Hugging Face cautions that recognition-based error rates act as a proxy for intelligibility, while embedding similarity estimates the preservation of voice identity. Neither measurement directly establishes naturalness, expressiveness or listener preference, so the automated rankings do not replace human preference testing.

Visitors can use a Listen tab to compare the generated outputs behind the metrics and vote on them. That gives people a way to judge the audio itself rather than relying only on leaderboard figures.

Fast first audio and small model files

Phonon’s timing results focus on how long users wait before hearing playable speech. Gradium’s chart of 95th-percentile CPU time to first audio placed Phonon 31m English first at 28 milliseconds and Phonon 40m English second at 38 milliseconds. Pocket TTS recorded 165 milliseconds in the same comparison.

“Lower values are better.”

“Time to first audio measures how long a user waits for playable speech, making it directly relevant to voice agents and other interactive applications.” A shorter wait matters when an application must respond as a conversation unfolds, especially when speech can begin before the full generation process ends.

The English Phonon 31m model card lists 64.5 MB of 8-bit quantized weights under Apache 2.0, a license that permits commercial use. The 40m variants have 74.6 MB of quantized weights and cover English, French, German, Spanish and Portuguese.

Those 40m languages are available through separate language checkpoints or one multilingual model, and they use the CC BY-NC 4.0 license. For commercial products built on those non-commercial checkpoints, Gradium prices licensing by the number of users rather than the amount of speech generated.

Voice options and deployment details

Phonon supports more than fixed voices. Gradium Studio can clone a voice from a short recording when the speaker gives consent, or design one from a text description for download and local use. That connects the speech models with both voice cloning and voice design while keeping the resulting use on the device.

The source material also includes the line: “Deployment Paths and Limits Gradium’s ptts runtime is written in Rust and does”. The statement ends there, leaving the available material without the rest of the sentence.

For now, the clearest picture is a compact speech system built around local processing, fast access to playable audio and multiple licensing paths. Its published measurements combine automated tests with a Listen tab, giving developers both performance numbers and a way to hear the outputs behind them.

Stay connected via Google News

Follow us for the latest AI News