phonon-31m English for audio.cpp

phonon-31m English for audio.cpp

Speaks English text aloud on a CPU from a 31M text-to-speech model in GGUF form, for the audio.cpp runtime. Made for developers who want a small local voice model.

Made by gradium on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is phonon-31m-en, an English text-to-speech model, converted to GGUF for the gradium_phonon family of audio.cpp. It is the same model, and only the file format differs. It needs an audio.cpp build that includes that family, and it does not load in llama.cpp.

Who it suits: It suits developers who want a small English voice model running locally through audio.cpp. Other Phonon runtimes cover the command line, Python, browsers, iOS and Android.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Software | Developers using audio.cpp | Add spoken output to a local tool | Read a status message aloud
Research | Speech researchers | Try a small text-to-speech model in GGUF | Compare it with other local voices
Accessibility | App builders | Speak interface text for users | Read on-screen text in a desktop app

If you want to run it in llama.cpp, you can skip it. The card names a separate repository for that.

Where to use it:

• audio.cpp, built with the gradium_phonon family
• A CPU, through the cpu backend
• Another audio.cpp backend, such as metal
• Streaming mode, when you want the audio to arrive as it is made

Title: Jobs it does
Use case | Who | What they get back
Speak a line of English | Developers | A WAV file of the speech
Stream speech | App builders | Audio delivered as it is generated
Switch the voice | Tinkerers | A different voice from a voice file you add

How to set it up:

• An audio.cpp checkout built with the gradium_phonon family, which the card says is not merged yet
• Python 3, to run the model manager script
• Disk space for a small model of about 21 MB at 4-bit by Hugging Face's estimate

1. Get an audio.cpp build that includes the gradium_phonon family.
2. Install the model with the model manager script.
3. Run the audiocpp_cli command to speak a line of text.

python3 tools/model_manager_v2.py install gradium_phonon_31m_en_q8_0 --models-root models

audiocpp_cli --task tts --family gradium_phonon --model models/phonon-31m-en \
  --backend cpu --text "Hello from Phonon." --out speech.wav

pip install -U huggingface_hub
hf download gradium/phonon-31m-en-audiocpp

How to use it: Run audiocpp_cli from the audio.cpp checkout with your text. Pick another backend with the backend flag, and add the mode flag set to streaming to stream the audio. Spell out numbers, symbols and abbreviations in words first.

Title: Options in the command
Tool | What it does
--backend | Picks the audio.cpp backend, such as cpu or metal
--mode streaming | Streams the audio as it is made
--voice-id | Chooses a voice you copied into the model folder
--text | The text to speak
--out | The WAV file to write

Example: You run the command with the text "Hello from Phonon." and the cpu backend. It writes a file called speech.wav with that sentence spoken in the default voice.

For newcomers and experts: If you are new, text-to-speech turns written words into a spoken voice. The first small thing to try is the install line and the one speak command from the card.

If you know the field, the single GGUF file holds the model, config and tokenizer. The transformer's linear layers are 8-bit, and the rest, including the Mimi decoder, keeps the original precision. The model builds on Pocket TTS and the Mimi codec from Kyutai. The trade-off is availability: the needed audio.cpp family is not merged.

Build audio.cpp with the Phonon family
Install the model files
Run audiocpp_cli with your text
A WAV file of speech is written

Good to know:

• Voices: without a voice id you get the default voice. Extra voices sit in the gradium/phonon-voices repository, and you copy one into the model folder. The model cannot clone a voice from a recording.
• Text: audio.cpp sends text to the tokenizer as written, so write out numbers, symbols and abbreviations.
• Sampling: temperature 0.3, the same as the Phonon runtime.
• License: apache-2.0.
• Cost to run: no API key, account or paid service is named.
• Data it reaches: you download files from Hugging Face, and the card names no other access.
• Not for: llama.cpp, which has its own repository, and cloning a voice from a recording.

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: gguf
• Parameters: 41.1M
• Weights: about 21 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 36,865
• Likes: 3
• License: apache-2.0
• Languages: en
• Based on: gradium/phonon-31m-en
• Tags: gguf, text-to-speech, phonon, audio.cpp

Links

• Model: https://huggingface.co/gradium/phonon-31m-en-audiocpp
• Files: https://huggingface.co/gradium/phonon-31m-en-audiocpp/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download gradium/phonon-31m-en-audiocpp

Then follow the usage in its model card. Nothing in this kit runs by itself.