Skip to content
trendoviz

Phonon launches offline TTS with two inference paths

Gradium has launched Phonon, a compact text-to-speech model designed to generate speech locally without a server or network connection. The developer reports strong leaderboard positions and provides separate routes for batched GPU synthesis through Transformers and single-stream device inference through its `ptts` framework.

27 out of 100Dormant7D: Insufficient history ProvisionalPublished

Overview

Phonon is a small text-to-speech model that runs on the user’s device, keeping the input text local and allowing speech generation to continue without a network connection [1]. Gradium reports two model variants: Phonon 31m supports English under Apache 2.0, including commercial use, while Phonon 40m covers English, French, German, Spanish and Portuguese under CC BY-NC [1].

The licensing split determines which variant can serve a commercial product. Gradium says Phonon 31m can be used in commercial products without restriction under Apache 2.0. For the other models, it says commercial pricing is based on the number of users rather than the amount of speech generated, but it has not published rates or contract terms [1].

Phonon uses flow matching to generate audio and is based on the Pocket TTS lineage [1]. The release includes more than 100 built-in voices, support for phone numbers, dates and codes, and custom-voice creation through voice cloning and voice design in Gradium Studio [1]. The developer also says the model handles text such as phone numbers, dates and codes [1].

Why it matters

Phonon gives developers two documented inference paths with different intended uses. Gradium says the Transformers implementation, built by the Hugging Face team and available from Transformers 5.21 as a reference PyTorch implementation, targets batched GPU generation and full-graph torch.compile for workloads such as synthesizing speech datasets at scale [1]. The developer’s ptts framework targets single-stream, on-device inference and supports more than 10 CPU, GPU and NPU backends, including Intel and Arm CPUs, WebAssembly, CUDA, Vulkan, Metal, WebGPU, Apple Core ML and Snapdragon [1].

The decision rule is direct: use Transformers for batched GPU synthesis and ptts for single-stream device inference [1]. Offline reading, accessibility tools, games and embedded interfaces are plausible applications for a model that needs no network connection, but the source does not document production deployments in those categories [1].

Opportunities

The open weights and reference implementation give teams two starting points: Transformers for integration into a PyTorch stack or study of the model, and ptts for deployment across heterogeneous hardware [1]. Gradium also reports more than 100 built-in voices and custom-voice creation through voice cloning and voice design in Gradium Studio [1].

The evidence describes several community and internal projects, but these should be treated as developer demonstrations rather than fully specified, ready-to-deploy components: an ExecuTorch conversion running Phonon on a visitor’s GPU in the browser, a Node.js integration called Micdrop, and the smallest-s2s Space [1]. Gradium reports more than 200,000 downloads in the first two weeks and continuing trending status on Hugging Face, though it does not provide independent confirmation of those figures [1].

Risks

Commercial use of the multilingual model depends on terms that are not set out in the release post. Phonon 40m covers five languages under CC BY-NC, and Gradium says commercial pricing for these models is based on user count rather than speech volume, but the rates and contract terms are not published [1].

Custom-voice features create a separate operational question. The release confirms that voice cloning, voice design and a speaker-similarity evaluation metric exist, but it provides no information about consent controls, identity verification, watermarking or anti-abuse measures [1].

Performance outside the reported leaderboard conditions is unknown. The source does not publish device-level latency, memory use or real-time-factor results for phones, laptops, single-board computers or browsers, so teams cannot infer production performance on those devices from the demonstrations [1].

Context

Cloud speech services require each input to be sent to a server and synthesized remotely. Phonon is designed to remove that network dependency by producing audio on the same device that holds the text [1]. The developer reports model sizes of about 65 MB and 75 MB and says Phonon runs on phones, laptops, single-board computers and in a browser tab, but it does not publish device-level latency, memory use or real-time-factor measurements for those systems [1].

Gradium reports that training was initially numerically unstable and required several techniques to produce consistent results across all five languages. That account describes the difficulty of training the multilingual model, not the quality of the released system at runtime [1].

Known, claimed, inferred

Gradium describes Phonon as having first place in German and Spanish for intelligibility, a top-three position in every language it evaluated alongside models up to 100 times its size, and the lead position in the CPU streaming time-to-first-audio benchmark [1]. These are developer-reported results, and the post provides no underlying scores, full ranking by language or complete test methodology [1].

The same distinction applies to the release information. Gradium says Phonon became available in Transformers 5.21 as a reference PyTorch implementation, built by the Hugging Face team. The availability of the integration is documented, but the release post does not include independent verification of its performance [1].

The leaderboard described by the developer evaluates 23 languages using word error rate for intelligibility, inverse real-time factor and time-to-first audio for speed, UTMOS for English and KOMOS for Korean for naturalness, and speaker similarity for voice cloning [1]. The term “high quality” therefore spans several distinct measurements rather than a single defined criterion [1].

Limitations & uncertainties

The account of Phonon relies on a first-party post from Gradium and includes no independent test report [1]. The roughly 65 MB and 75 MB figures are not explicitly mapped to the 31m and 40m variants in the source [1].

The leaderboard rankings, evaluated language list beyond the stated top-three result, and device behavior remain unverified because the underlying table and scores are not included [1]. No information is provided about the quality of outputs from the voice-cloning or voice-design workflows beyond their availability [1]. The technical explanation in the source is truncated, so the complete architecture and training description is unavailable [1].

Evidence & sources

The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.

  • Primary-source or vendor statementGradium launched Phonon, a text-to-speech model that runs locally without a server or network connection and keeps typed text on the device.[1]
  • Primary-source or vendor statementGradium reports that Phonon 31m supports English under Apache 2.0 and is usable in commercial products, while Phonon 40m covers five languages under a non-commercial license.[1]
  • Primary-source or vendor statementGradium reports that Phonon ranked first in German and Spanish for intelligibility, placed in the top three in every language it evaluated, and led the CPU streaming time-to-first-audio benchmark; no underlying scores are published.[1]
  • Primary-source or vendor statementGradium says Phonon became available in Transformers 5.21 as a reference PyTorch implementation, built by the Hugging Face team.[1]
  • Primary-source or vendor statementThe documented inference choices are Transformers for batched GPU synthesis and the developer's `ptts` framework for single-stream device inference across more than 10 CPU, GPU and NPU backends.[1]
  • Primary-source or vendor statementGradium reports more than 100 built-in voices, custom-voice creation through voice cloning and voice design in Gradium Studio, and handling of phone numbers, dates and codes.[1]

Sources

  1. 1

Updates & corrections

  1. First published · published automatically after passing every check

Found an error? Report it.