phonon-31m English for llama.cpp

phonon-31m English for llama.cpp

Turns English text into speech from the command line or a local server, as a GGUF copy of the Phonon 31M model for llama.cpp. It suits developers who already run llama.cpp with a Phonon-enabled build.

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 the GGUF conversion of the Phonon 31M English model, made for the llama-tts and llama-server programs of llama.cpp. It reads English text and writes speech. The model is the same, and only the file format differs.

It works only with a llama.cpp build that has Phonon support, and that support has not been merged into the main project yet.

Who it suits: Developers who already run llama.cpp and want local English speech from text suit it best. No cloud service is involved.

Industry | Who uses it | How they use it | Example
Software | Developers | Add local text to speech to an app through the llama.cpp server | Send text to the server and save the audio
Voice tools | Hobbyists | Make speech files from the command line | Turn one sentence into a wav file
Research | Audio researchers | Test a small speech model inside the llama.cpp ecosystem | Compare it with the Phonon runtime

Anyone who wants voice cloning, another language, or Phonon on a phone or in a browser can skip it and use the main Phonon model.

Where to use it: It runs through two llama.cpp programs, on any computer where you have a Phonon-ready llama.cpp build.

• llama-tts, for speech files from the command line
• llama-server, for a local server that returns audio
• A llama.cpp build with Phonon support
• Windows, macOS or Linux with Python for the file download

Use case | Who | What they get back
Speech files | Developers | A wav file for a line of English text
Local speech server | Developers | Audio returned for each text request
Offline voice features | Hobbyists | Speech made on your own machine

How to set it up: You need a build of llama.cpp that includes Phonon, plus the two model files. The record estimates about 11 MB at 4-bit, so disk space is small.

• A llama.cpp build that has Phonon support
• Python and the huggingface_hub package, if you download the files yourself
• Both model files, because llama.cpp needs the two together

File | What it holds
Model file | The main speech transformer in 8-bit, plus the tokenizer and a built-in voice
Projector file | The Mimi decoder and the flow head, kept at full precision

1. Get a llama.cpp build with Phonon support.
2. Install the hub tools with pip.
3. Download the model files with the hf command.
4. Run llama-tts or llama-server with the two files.

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

How to use it: The model file goes in with the -m flag, and the projector file goes in with -mm. The first command below lets llama-tts fetch the files and speak one line.

llama-tts -hf gradium/phonon-31m-en-llamacpp -p "Hello from Phonon." -o speech.wav

With the files already downloaded, you name both of them yourself.

llama-tts -m phonon-31m-en-Q8_0.gguf -mm mmproj-phonon-31m-en-F32.gguf \
  -p "Hello from Phonon." -o speech.wav

To run it as a server, start llama-server on a port. Then send a POST request to its tts path with a JSON body that holds your input text, and save the reply as a wav file.

llama-server -hf gradium/phonon-31m-en-llamacpp --port 8080

Example: You run the first command with a short English sentence in the -p flag. It writes a wav file with that sentence spoken aloud.

For newcomers and experts: Think of this as a tiny voice that reads English text aloud on your own computer. Your first small try is to run the first command with a short sentence, once you have a Phonon-ready llama.cpp build.

Under the hood, the main file holds a flow-matching transformer in 8-bit, and the projector file holds the flow head and the Mimi decoder. Phonon builds on Pocket TTS and the Mimi codec from Kyutai. The trade-off is limited voices and a build that is not yet in the main project.

Your English text
llama.cpp tokenizer
Phonon transformer and flow head
Mimi decoder makes the audio
Speech saved as a wav file

Good to know: Several limits come from how llama.cpp handles text and voices today.

• Voice: a default voice is built in, so you need no speaker file, and there is no voice cloning from audio
• Other voices: the extra voices in the Phonon voices collection do not work in llama.cpp yet
• Text: llama.cpp passes your text to the tokenizer as written, so spell out numbers, symbols and short forms as words
• Phonon runtime: it spells these out for you when you use --lang en
• Characters: emoji and letters the tokenizer lacks reach the model as an unknown token, so use plain text
• Sampling: temperature 0.3, the same default as the Phonon runtime
• Size: 21M parameters in the record, in GGUF format, English only
• License: apache-2.0
• Cost to run: none named
• Data it reaches: the text you give it, and the files downloaded from Hugging Face

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: gguf
• Parameters: 21M
• Weights: about 11 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 348
• Likes: 3
• License: apache-2.0
• Languages: en
• Based on: gradium/phonon-31m-en
• Tags: gguf, text-to-speech, phonon, llama.cpp

Links

• Model: https://huggingface.co/gradium/phonon-31m-en-llamacpp
• Files: https://huggingface.co/gradium/phonon-31m-en-llamacpp/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-llamacpp

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