Phonon TTS: licenses, local voices and runtime choices
Phonon TTS gives developers a small speech model to run on local hardware, but choosing a checkpoint involves more than comparing download sizes. The important decisions are commercial licensing, the inference runtime and how custom voices reach the device.
Gradium released the open weights on October 20, 2026. Its November 5 technical follow-up examines how to use them across different workloads, from a single on-device stream to batched GPU generation. November 5 technical follow-up.
Check the exact model license
The English phonon-31m-en model card lists Apache 2.0 for its weights, tokenizer and included voices, allowing commercial use. Its 8-bit weights file is 64.5 MB. The 31-million-parameter figure excludes the audio tokenizer, so it should not be treated as a complete count of every component needed for synthesis. Output is mono speech at 24 kHz. English model card.
The multilingual 40M checkpoint covers English, French, German, Spanish and Portuguese in a 74.6 MB quantized weights file. Its weights, tokenizer and included voices use CC BY-NC 4.0, which does not permit commercial use. Its family table gives the same license for the five language-specific 40M checkpoints. A commercial product using those models needs separate permission; downloadable weights alone do not establish that right. Multilingual model card.
For an English-only prototype intended for commercial deployment, begin the license review with the exact 31m-en repository. For a multilingual product, resolve licensing before investing in integration. Keep the model, runtime and any additional voice assets separate in the dependency inventory.
Choose the runtime around the workload
The November follow-up describes ptts as the single-stream, on-device route. Transformers supplies the reference PyTorch implementation, with batched generation and full-graph compilation aimed at GPU throughput. An interactive assistant and a job producing thousands of speech files therefore have different starting points. November 5 technical follow-up.
The xn-ptts runtime is written in Rust and does not require PyTorch. Its documentation provides desktop, Python, browser and mobile paths. Default configurations download and cache model files on first use. Teams planning offline operation should package or preload the required files, then test startup without a connection rather than assuming the first run is offline. Phonon runtime documentation.
Custom voice preparation is a separate dependency
Transformers documentation says Phonon uses precomputed speaker latents because the speaker codec has not been released. The distributed model cannot turn a fresh recording into a voice prompt by itself. Transformers documentation.
The runtime instead documents designing or cloning a voice through Gradium Studio, then downloading a voice file for local synthesis. Its command-line preparation scripts require a Gradium API key, and cloned voices are also saved in the Gradium account. Obtain the speaker’s consent and review this preparation step separately from the eventual local playback path. Phonon runtime documentation.
Test pronunciation and sustained listening
Transformers also differs in text preparation: its tokenizer does not expand language-dependent numbers, monetary amounts, symbols or addresses. The documentation recommends Gradium’s normalization or writing those forms out. Include product codes, dates and prices in acceptance tests. Transformers documentation.
The English model card warns that long inputs are split into sentences, so prosody does not carry across an entire text. Listen to paragraphs and transitions, not just a clean opening sentence, before choosing it for narration. English model card.
Read device numbers with their conditions
Gradium’s launch tests report roughly 7 ms to first audio on a MacBook Air M5 CPU and 38 ms on a Raspberry Pi 5. These are vendor measurements. The protocol uses one English sentence, warm-up runs and median timings; first-audio latency excludes one-off voice conditioning. Browser measurements use a different sentence and best runs. Gradium launch and test protocol.
Use those results to select hardware for a trial, then measure cold startup, peak memory, audio continuity and longer sessions in the actual application. A small weights file and a fast prepared example are useful signals, but neither answers every deployment question.
For a broader local assistant with audio and other inputs, see our guide to Liquid AI’s LFM3 models. Choosing a speech synthesizer is one part of designing the complete application.
