How Much To Run AI
Open weights

Phonon: On-Device TTS, Variant Licenses and Running Cost

Phonon is Gradium’s family of streaming text-to-speech models for local applications. It turns text into mono 24kHz speech, including the audio decoding needed to produce a waveform. The 31M English model is Apache 2.0; the 40M single-language and multilingual models use CC BY-NC 4.0. Choose the variant and its use rights before comparing device costs.

Content reviewed:

Use this model

Official model repository

Local tools and guides

Use cases

  • Generate spoken English from a fixed script on a local device.
  • Compare a single-language 40M checkpoint with the five-language variant.
  • Add streaming speech to an application using the matching runtime.

Choose a language and license

Phonon 31M English is the public Apache 2.0 option. Each 40M single-language checkpoint covers English, French, German, Spanish or Portuguese; the 40M multilingual checkpoint covers those five languages in one model. Both 40M routes are CC BY-NC 4.0, so commercial use requires separate permission from Gradium. The runtime’s MIT OR Apache-2.0 license does not replace the model terms.

The 40M repositories require accepting their Hub conditions before file access; the runtime documents automatic approval. Server and desktop clients can use a saved Hub login. A browser cannot supply that login through the public model loader, so 40M browser integration requires serving the authorized files with your application under their license.

VariantLanguage scopeModel terms / q8 file
phonon-31m-enEnglishApache 2.0 · 64.5MB model.q8.gguf
phonon-40m-en/fr/de/es/ptOne language per checkpointCC BY-NC 4.0 · 74.6MB model.q8.gguf
phonon-40m-multilingualEnglish, French, German, Spanish, PortugueseCC BY-NC 4.0 · 74.6MB model.q8.gguf

Generate speech with the official runtime

Use Gradium’s xn-ptts runtime for the complete text-to-waveform path. It provides desktop, Python, browser, Swift and Android integrations. The Python example below selects the public English checkpoint, q8 weights and English normalization. Its first load downloads model files; later loads can reuse the cache. Listen to the saved WAV and check numbers, names and punctuation on a script that resembles your application.

The model predicts audio latents and the Mimi decoder turns them into sound. This is a complete TTS workflow, not an audio encoder that can be priced on its own. Reuse the loaded runtime for a batch rather than including another cold load for every sentence.

After installing the ptts Python package; downloads files on first use
import ptts

tts = ptts.TTS("hf://gradium/phonon-31m-en", lang="en", quant="q8")
tts.save("speech.wav", "Your appointment is at ten thirty.")

Account for the tokenizer, decoder and application

The published 31M and 40M parameter counts exclude the audio tokenizer. The 64.5MB and 74.6MB q8 figures identify weight files, not full process RAM. Include the codec, voice files, working buffers and application when measuring memory. A browser also has its own runtime and caching overhead.

Choose the backend supported by your target device and measure cold loading separately from warm synthesis. Record elapsed generation time per minute of finished speech, peak process memory, the exact model revision and voice. Parameter count and download size cannot supply a portable real-time factor or a GPU memory minimum.

Compare local speech using measured time

A local batch budget starts with finished audio hours × measured generation minutes per output minute. For an assumed two hours of finished speech at RTF 0.5, generation occupies one device-hour. At an illustrative whole-system 20W and $0.20/kWh, that hour uses $0.004 of electricity, before loading, idle time and hardware allocation. This is a formula example, not a Phonon benchmark.

The shared speech calculator estimates costs for a custom GPU-backed TTS setup from your measured full-rig time, peak memory and rates. CPU, phone and browser deployments need separate device-cost accounting. It has no Phonon speed preset, minimum VRAM or hosted API price. Character-billed API comparisons use a separate workload and do not convert characters into audio minutes.

Before you use it

  • Commercial rights differ across Phonon variants; the small parameter count does not describe the memory of the complete TTS process.

How to get started

Choose the permitted checkpoint and language, generate a representative script with the official runtime, then measure the complete device session and usable audio output.

Sources

Sources (4)

Calculators