Gradium 端侧 TTS 模型 Phonon 两周下载超 20 万次

Gradium reported on November 5, 2026 that its Phonon on-device text-to-speech models had been downloaded more than 200,000 times in their first two weeks, in a technical follow-up published on Hugging Face.

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

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, with Pocket TTS at 165 milliseconds. 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.

What the Rankings Measure

The follow-up describes evaluation across 23 languages covering intelligibility, inverse real-time factor, time to first audio and automated naturalness measures, with speaker similarity added for voice cloning. Hugging Face’s leaderboard methodology calculates intelligibility by comparing automatic speech recognition transcripts with the original prompts using word or character error rates.

Streaming tests run one utterance at a time using common prompts and the same hardware. For models without a streaming interface, the timing covers generation of the entire utterance because playback cannot begin earlier. Voice-cloning evaluations separately compare speaker embeddings from generated audio and a reference clip.

Hugging Face cautions that recognition-based error rates are a proxy for intelligibility, while embedding similarity estimates preservation of voice identity. Neither measurement directly establishes naturalness, expressiveness or listener preference, so the automated rankings do not replace human preference testing. A Listen tab allows visitors to compare the generated outputs behind the metrics and vote on them.

Open Weights and Commercial Terms

The underlying open-weight release came on October 20, 2026, making the new disclosure an early adoption and evaluation update rather than a separate model debut. The English Phonon 31m model card lists 64.5 MB of 8-bit quantized weights under Apache 2.0, which permits commercial use.

The 40m variants have 74.6 MB of quantized weights and cover English, French, German, Spanish and Portuguese under CC BY-NC 4.0, through separate language checkpoints or one multilingual model. For commercial products built on those non-commercial checkpoints, Gradium prices licensing by the number of users rather than the amount of speech generated.

Phonon produces mono 24 kHz audio and can stream speech while generation continues. It is intended for offline assistants, accessibility tools, games and interactive applications that cannot wait for a cloud round trip, while keeping text on the user’s device. More than 100 built-in voices are available. Gradium Studio can also clone a voice from a short recording with the speaker’s consent or design one from a text description for download and local use.

Deployment Paths and Limits

Gradium’s ptts runtime is written in Rust and does not require PyTorch. It supports CPU, GPU and NPU backends spanning Intel, Arm, WebAssembly, CUDA, Vulkan, Metal, WebGPU, Apple Core ML and supported Snapdragon hardware. Its packages cover browser, mobile, desktop and local-server integrations.

Phonon is also available in Transformers from version 5.21 as a reference PyTorch implementation. While ptts focuses on a single on-device stream, Transformers targets GPU throughput through batched generation and full-graph torch.compile. The Transformers documentation describes a causal transformer producing continuous 32-dimensional audio latents at 12.5 Hz for decoding into 24 kHz audio. Voices are supplied as 512-dimensional precomputed speaker latents.

Those mechanics leave some integration work outside the checkpoint. The speaker codec is unreleased and the audio tokenizer is decoder-only, so Transformers cannot clone voices directly from recordings. Its tokenizer also does not expand numbers, amounts, symbols or addresses as Gradium’s runtime does. Long inputs are split into sentences, preventing prosody from carrying across an entire text.

The deployment choice is therefore concrete: ptts serves a private, single stream on the user’s hardware, while Transformers serves batched GPU synthesis. In both paths, voice preparation and language-aware handling of numbers, symbols and addresses remain part of the integration alongside the checkpoint itself.