A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That

The announcement beats are fixed. I'll lock paragraph length to the source facts, then check the deck length.Gradium opened Phonon, its small text-to-speech model, two weeks before a November 5, 2026 post by Olesia Krizhevich (olesiagradium), Laurent Mazare (lmz), Eric Bezzam (bezzam), and Eustache Le Bihan (eustlb). Hugging Face Blog's report says Phonon runs on the device with no server and no network, and that typed text stays there.

The post covers why Gradium built an on-device voice, what Phonon can speak, the demos, and results on the Open TTS Leaderboard. It is for developers putting speech on a phone, laptop, single-board computer, or browser tab, and for teams choosing Apache 2.0 or CC BY-NC.

A Closer Look at Phonon TTS: the announcement

Most voice AI lives in the cloud, and every sentence makes a round trip to a server. Phonon keeps text on the device and speaks with no network. Gradium opened the weights. Phonon 31m is Apache 2.0, free in a commercial product. Other models are CC BY-NC, free for non-commercial work, with commercial pricing by number of users, not speech generated. Announced in April, Phonon then had six months with partners and early users. Gradium has been "very happy with the reception so far."

That feedback made the models smaller and added languages and custom voices through voice cloning and voice design. Two sizes, about 65 MB and 75 MB, together speak English, French, German, Spanish, and Portuguese. More than 100 built-in voices are included, and Gradium Studio can create more. Phonon reads phone numbers, dates, and codes, where many TTS systems stumble. It runs on phones, laptops, single-board computers, and in a browser tab. Phonon 31m is English under Apache 2.0. Phonon 40m covers all five languages under a non-commercial license.

What actually changed with A Closer Look at Phonon TTS

Phonon comes from the Pocket TTS lineage and uses flow matching, unlike other TTS modeling. That put this quality in a small model, but training was numerically unstable and needed several tricks for all five languages. A small model did not sound great, and a strong voice would not fit. Gradium says this release addressed both. ptts, a new inference framework, supports more than 10 backends: CPU on Intel, Arm, and WebAssembly; GPU on CUDA, Vulkan, Metal, and WebGPU; and NPU on Apple Core ML and Snapdragon.

Transformers ships Phonon from v5.21 as its reference PyTorch implementation. ptts runs one stream on device. Transformers targets GPU throughput: batched generation, full-graph torch.compile, and speech datasets at scale. Native PyTorch is a base for another stack or for study. Pocket TTS was added first. Phonon adds about 200 lines for what differs. Audio latents are generated autoregressively, each fed back as the next input, with the loss computed on them. Latents are the inputs, outputs, and targets, as text tokens are for a language model, so the codec is the tokenizer. The decoder-only tokenizer sits beside the text tokenizer. Phonon never encodes audio with it.

Who should care about A Closer Look at Phonon TTS

An Apple Watch with Wi-Fi off uses Phonon to talk through a coffee brew step by step. An iPhone with no signal counts reps. Lamp, a robot from Autonomous, wears a sorting hat and talks in character. A browser turns an offline dino runner into a radio show hosted by Rex. A sci-fi game is voiced on the player's own machine, and Proust is read page by page. English commercial apps can use Phonon 31m under Apache 2.0. Phonon 40m, non-commercial, speaks English, French, German, Spanish, and Portuguese.

The Hugging Face team built Transformers support so Phonon worked there from day one. ExecuTorch runs it on a visitor's GPU in the browser. Micdrop brings it to Node.js apps. Spaces such as smallest-s2s build on it. Downloads passed 200,000 in the first two weeks, and Phonon was still trending. Gradium wrote, "we love it," about the Open TTS Leaderboard, a new Hugging Face project where more than 9,000 TTS models on the Hub can be compared and heard side by side. The post lists three models, one dataset, and three Spaces.

How to try A Closer Look at Phonon TTS

Across 23 languages, scores cover intelligibility (word error rate), speed (inverse real-time factor and time-to-first audio), and naturalness (UTMOS for English and KOMOS for Korean). Fixed-voice and voice cloning are both evaluated, with a speaker-similarity metric for cloning. A Listen tab compares outputs side by side so a visitor can vote. Transformers v5.21 loads gradium/phonon-31m-en with AutoProcessor and GradiumPhononTTSForConditionalGeneration, device_map set to auto. The model reports 30,788,354 parameters. The audio tokenizer reports 10,305,313. A voice is a tensor of precomputed speaker latents, not a recording.

Those latents live in gradium/phonon-voices. hf_hub_download fetches voices/LMDIYE0Oa377Qh6Z.safetensors, and safetensors load_file reads speaker_wavs. Transposed, the first tensor has shape (num_frames, 512). The chat template orders voice positions, text, then a token that opens the audio, and the same conversation works across checkpoints. A user turn holds an empty audio item and "Hello from Phonon." apply_chat_template uses tokenize, return_dict, add_generation_prompt, and audio_prompt_latents. The processor expands that placeholder to one position per frame, then normalizes, tokenizes, and batches. The prompt repeats <|audio_prompt|> for every frame, then the sentence, then <|audio_bos|>. generate returns audio_latents shaped (batch_size, num_frames, 32). processor.decode and save_audio write phonon.wav.

What to watch after A Closer Look at Phonon TTS

Phonon ranks first on intelligibility in German and Spanish, and in the top three in every language Gradium evaluated, beside models up to 100 times its size. The models top the CPU streaming benchmark on time-to-first audio. A companion article dated September 30, 2026 is titled Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning. Another voice can be cloned from a short recording with the speaker's consent, or described in plain text in Gradium Studio, then downloaded and run on the device.

Gradium trains custom Phonon models with design partners. Tight memory can mean a smaller model. Korean, Arabic, or Japanese can be trained on request. A product vocabulary can be fine-tuned. The post says to get in touch and state the need. It thanks the Hugging Face team and everyone who tried Phonon in its first two weeks. Gradium calls Phonon its first open release and asks builders to say what they make. Commercial CC BY-NC terms follow the number of users, not speech generated.