Key takeaways
- Hugging Face published a detailed Phonon review on November 5, 2026. Gradium originally released the model family on October 20, so this is fresh technical and leaderboard coverage of an available release—not a new model appearing for the first time today.
- The exact English checkpoint is gradium/phonon-31m-en: 31M model parameters excluding the audio tokenizer, 64.5 MB 8-bit weights, 24 kHz mono speech and an Apache 2.0 license. Its repository was public and ungated during review.
- The 40M family includes language-specific and multilingual checkpoints. The multilingual model covers English, French, German, Spanish and Portuguese in 74.6 MB 8-bit weights, but it is CC BY-NC 4.0 and its files require accepting an access gate.
- Gradium says Phonon ranks first for intelligibility in German and Spanish and leads the CPU-streaming time-to-first-audio benchmark. Those results come from automated leaderboard metrics and vendor-associated reporting, not an independent production listening study.
- Prototype with the Apache-licensed 31M English model when its language and quality fit. Treat multilingual commercial use as a separate licensing and procurement decision, and evaluate pronunciation, voice consent, latency and device behavior on representative hardware.
Hugging Face published the technical review on November 5
Hugging Face and Gradium published a closer look at Phonon on November 5, 2026, covering the release, demos, leaderboard placement and Transformers integration. The weights themselves are not brand-new: Gradium’s launch article is dated October 20, and the Hugging Face registry records the two principal repositories as created October 19. The current event is a detailed follow-up that makes the model family and its evaluation evidence easier for builders to inspect.
Phonon targets local, streaming text-to-speech rather than a metered cloud endpoint. Gradium says it runs across phones, laptops, single-board computers and browsers. Its ptts runtime lists CPU, GPU and NPU backends including Intel, Arm, WebAssembly, CUDA, Vulkan, Metal, WebGPU, Apple Core ML and Snapdragon. Transformers provides a separate reference PyTorch path for batched GPU generation from version 5.21.
The 31M English checkpoint has the simplest commercial path
gradium/phonon-31m-en is a public, ungated repository with safetensors and GGUF artifacts. Its model card specifies 30.8 million model parameters excluding the roughly 10.3 million-parameter audio tokenizer, 64.5 MB 8-bit weights and 24 kHz mono output. It streams one 80 ms audio-latent frame at a time before the Mimi-derived decoder turns latents into audio.
The repository is tagged Apache 2.0, which permits commercial use subject to that license’s terms. That makes this checkpoint the clearest starting point for an English-only product prototype. “Open weight” still does not establish acceptable pronunciation, naturalness, device thermals, battery use or end-to-end latency; those require measurements on the intended hardware and content.
The 40M models change both language coverage and licensing
The 40M family has English, French, German, Spanish and Portuguese language-specific checkpoints plus gradium/phonon-40m-multilingual, which covers all five languages in one set of weights. The model card lists 74.6 MB for its 8-bit weights and the same 24 kHz mono output. During review, the repository page was public but required a signed-in user to accept conditions before accessing files.
These checkpoints use CC BY-NC 4.0 rather than Apache 2.0. Gradium says commercial terms are based on user count rather than generated speech, but neither reviewed launch page publishes a price table. A free download for non-commercial work is therefore not evidence of a free commercial deployment. Buyers need written terms for their use case before shipping the 40M variants.
The benchmark lead is useful evidence, not a listening verdict
The November 5 article says Phonon ranks first on intelligibility in German and Spanish, places in the top three in every evaluated language and leads the CPU-streaming benchmark for time to first audio. Gradium separately reports a 7 ms start for Phonon 31M on a laptop CPU. AccessAllGPT did not reproduce those results, and Gradium authors contributed to the Hugging Face article reporting them.
The Open TTS Leaderboard evaluates intelligibility with word error rate, speed with inverse real-time factor and time to first audio, and naturalness with UTMOS for English or KOMOS for Korean. Voice-cloning entries also receive a speaker-similarity metric. These automated measures improve comparability, but they do not fully capture listener preference, accent fidelity, expressive control, difficult names, domain vocabulary or failures on long and malformed text.
Voice files are a separate consent and governance surface
Phonon represents a voice as precomputed speaker latents rather than a recording passed into each generation request. Gradium offers more than 100 built-in voices and says users can clone a consenting speaker or design a voice in Gradium Studio, then download the voice file for local use. Local inference can reduce routine text transmission, but it does not make the creation, storage or distribution of a cloned voice risk-free.
Teams should record who supplied each voice, the consent scope, permitted products, retention period, revocation path and whether synthetic output must be disclosed. Keep voice artifacts access-controlled and versioned separately from model weights. Test misuse controls and deletion workflows before making a custom voice available to end users.
What builders and buyers should do next
For an English prototype, begin with gradium/phonon-31m-en at a pinned repository revision and the runtime that matches the target device. Build a listening set containing names, dates, phone numbers, abbreviations, codes, punctuation, long passages and adversarial text. Measure time to first audio, real-time factor, peak memory, package size, battery or power use, crashes and listener acceptance on the slowest supported hardware.
Use a 40M checkpoint only after the license and access conditions fit the project. Compare language-specific and multilingual variants on native-speaker review rather than assuming one checkpoint wins everywhere. Trial when offline speech, privacy or recurring cloud cost matters; constrain custom voices behind consent and access controls; wait when commercial rights are unresolved; reject the deployment if representative pronunciation or device performance misses the product threshold.
Copy-ready Phonon deployment trial record
Keep model quality, device fit, licensing and voice governance in one reviewable record.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Product surface, users, English/French/German/Spanish/Portuguese needs, offline requirement and failure consequence.
Exact repository, revision, 31M or 40M, language variant, license, accepted gate terms and commercial approval.
ptts or Transformers version, backend, operating system, CPU/GPU/NPU, memory, package size and lowest supported hardware.
Representative prose, names, dates, numbers, codes, abbreviations, domain terms, long text, malformed input and native-speaker samples.
Listener acceptance, pronunciation errors, time to first audio, real-time factor, peak memory, power, crashes and recovery.
Voice source, consent, permitted uses, disclosure, storage, access, retention, revocation and deletion test.
Artifact verification, sandboxing, input limits, monitoring, fallback voice or service, rollback and incident owner.
Trial, constrain, wait or reject; approved model, devices, languages, traffic scope, unresolved evidence and retest date.
Primary sources
Browse the publication-wide evidence index →
- A Closer Look at Phonon TTSHugging Face and Gradium · Reviewed: Publication date; release summary; model sizes and languages; licenses; Open TTS Leaderboard claims; inference runtimes; Transformers integration; voice workflow · Retrieved · Supports: The November 5 article documents the Phonon family, exact Transformers model ID, supported runtimes and Hugging Face leaderboard claims. Gradium authors participated, so benchmark characterizations remain vendor-associated rather than independent validation.
- Open-Sourcing Phonon: A Suite of Frontier On-Device TTS ModelsGradium · Reviewed: October 20 announcement; model family; claimed latency; language coverage; platform support; evaluation charts; licensing and commercial access · Retrieved · Supports: Gradium announced the weights on October 20, identifies the 31M and 40M variants, reports a 7 ms start on a laptop CPU for 31M and describes the runtime targets. Performance and quality statements are vendor claims.
- gradium/phonon-31m-en model card and repositoryGradium on Hugging Face · Reviewed: Repository metadata; model details; family table; files; audio format; license; runtime and usage links · Retrieved · Supports: The public, ungated repository identifies a 31M-parameter English TTS model, 64.5 MB 8-bit weights, 24 kHz mono output and Apache 2.0 licensing, with safetensors and GGUF artifacts.
- gradium/phonon-40m-multilingual model card and repositoryGradium on Hugging Face · Reviewed: Access gate; model details; family table; languages; files; audio format and license · Retrieved · Supports: The public but access-gated repository identifies a 40M-parameter multilingual model, 74.6 MB 8-bit weights, five languages, 24 kHz mono output and a CC BY-NC 4.0 license.
- Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice CloningHugging Face · Reviewed: Leaderboard scope; languages; fixed-voice and voice-cloning tracks; intelligibility, naturalness, similarity and speed metrics; limitations · Retrieved · Supports: Hugging Face describes the leaderboard methodology and its automated metrics. This supports what the reported ranks measure, not production quality or listener preference for a specific application.
- Gradium Phonon TTS documentationHugging Face Transformers · Reviewed: Transformers version; architecture overview; processor and model classes; generation and decoding workflow · Retrieved · Supports: The live Transformers documentation confirms library support and the reference PyTorch integration surface.
Limitations
This is a first-party documentation, registry and methodology review, not a hands-on audio evaluation, independent benchmark, security audit, privacy assessment, legal opinion or procurement review. AccessAllGPT did not download or hash the weights; install ptts or Transformers; synthesize, listen to or compare speech; test a built-in or cloned voice; inspect training data; validate the claimed 7 ms start; reproduce leaderboard ranks; measure intelligibility, naturalness, similarity, latency, throughput, memory, power, package size or stability; test CPU, GPU, NPU, browser or mobile backends; review the full multilingual access-gate agreement while authenticated; obtain commercial pricing; or determine whether a particular use satisfies Apache 2.0, CC BY-NC 4.0 or voice-rights law. Performance, download and quality statements are time-sensitive vendor or platform claims.
Disclosures
AccessAllGPT received no weights, hosted access, commercial license, private demo, briefing, benchmark data, voice files, payment or compensation for this article. Gradium and Hugging Face did not sponsor, review or endorse it. AccessAllGPT did not test, score, rank or certify Phonon. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Gradium or Hugging Face. Publication-wide relationships are listed on the disclosures page.
