Mach Boom

Mach Boom

Turns English text into speech on a single CPU core, using a 3.5-bit version of Gradium's Phonon 31m model that runs several times faster than real time, for developers who need small, local voice output.

Made by SyzygyResearch 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: Mach Boom is a text-to-speech model for English. It squeezes Gradium's Phonon 31m model down to 3.5 bits per weight, which leaves a file of under 20 MB. On a single CPU core it produces audio roughly eight to ten times quicker than playback, per the card, and it needs no GPU.

Who it suits: Developers who want spoken output from a small local model will find it fits, especially on machines with only a CPU. Four voices come with it, and more voice files can be added.

Industry | Who uses it | How they use it | Example
Software | App developers | Add spoken output to a program without a GPU or a hosted service | Reading status messages aloud
Accessibility | Tool builders | Read written text aloud on ordinary CPU machines | A screen reading helper
Bookings and travel | Service developers | Speak confirmations that hold codes, times and dates | A spoken booking reference
Edge computing | Embedded developers | Run speech on small, low-memory machines | Voice output on a modest server

If you work on Windows or macOS, you can skip it, since the card lists Linux on x86-64 as the requirement.

Where to use it:

• Linux on x86-64, from the command line
• Python scripts, through the MachBoom class
• A pipe into an audio player, for streaming playback

Use case | Who | What they get back
Make a WAV file | Anyone with a text | A spoken file in the voice you pick
Stream speech | App developers | Audio pieces as soon as the first chunk is ready
Tune threads | Performance engineers | Faster speech with more CPU threads

How to set it up: You need Linux on x86-64 and a C compiler, since the first run builds the speech kernels for your CPU in about 5 seconds.

• Linux on x86-64
• Python 3.8 or newer
• numpy and tokenizers
• gcc or clang

1. Install the Hugging Face hub package.
2. Download the model into a folder named mach-boom and move into it.
3. Install numpy and tokenizers.

pip install -U huggingface_hub
hf download SyzygyResearch/Mach-Boom --local-dir mach-boom && cd mach-boom
pip install numpy tokenizers

How to use it: Run tts.py with your text, an output file and a voice name. Add threads to speed it up, or stream raw audio to a player.

python tts.py "Your booking reference is AB-4471-Z, and the meeting moved to 3:30 p.m." -o out.wav --voice Zoey
python tts.py "..." -o out.wav --threads 4
python tts.py "..." --pcm --first 2 | aplay -q -r 24000 -f S16_LE

From inside the mach-boom folder, the same thing works in Python.

from mach import MachBoom

tts = MachBoom("mach-boom.bin", voices_dir="voices", tokenizer="tokenizer.json")
wav = tts.synth("Call 020 7946 0018 before 5:30 p.m. on 14/03.", "Zoey", seed=0)  # float32, 24 kHz
for piece in tts.synth_stream("Your table is booked for 8 p.m.", "Zoey", first=2):
    ...

Tool | What it does
tts.py | Command-line script that writes a WAV file or streams raw audio
MachBoom class | Python interface with synth for a whole clip and synth_stream for pieces
voices folder | Holds the four shipped voices, Zoey, Juniper, Everly and Harper

Example: You give tts.py a sentence with a booking code and a meeting time, pick the voice Zoey and name out.wav as the output. The file comes back with that sentence spoken, numbers and times included.

For newcomers and experts: Text to speech reads written words aloud. A good first step is to download the folder, install the two packages and run the one-line command with a sentence of your own.

The code builds for several CPU instruction sets, from the newest AVX-512 forms down to plain C. Newer Xeons with AMX also decode audio in bf16 tiles. Fix the seed and the chunk size, and the audio stays identical whatever thread count you pick. In its own tests on one core, peak memory was 93 MB, and first audio came in 61 to 78 ms when streaming. The scores come from the authors, on their own hard test set.

Your text and a voice name go in
Kernels compile for your CPU on the first run
The 3.5-bit model generates the speech
A 24 kHz WAV file or audio stream comes out

Title: Word error rate on hard English sentences, lower is better
Mach Boom: 5.18 percent
Underdog Woof Juno: 6.50 percent
Fish Audio Drama 3: 7.62 percent
Phonon 31m fp32: 8.61 percent
Phonon 31m 4-bit: 10.61 percent
Kokoro 82M: 14.07 percent
Pocket TTS: 18.13 percent
Kitten Nano 0.8: 18.56 percent

Good to know:

• License: Apache-2.0, shared with Phonon 31m and its voices
• Cost to run: none named
• Data it reaches: it reads the text you give it and writes audio files, and the source names no web or account access
• Four voices ship with it, and voice files from Gradium's Phonon voices set drop in
• Not for: other languages or systems, since the card lists English and Linux on x86-64
• Setting PHM_AMX=0 turns off the AMX path on newer Xeons

On Hugging Face, 13 November 2026

• Task: text to speech
• Downloads in the last 30 days: 0
• Likes: 23
• License: apache-2.0
• Languages: en
• Based on: gradium/phonon-31m-en
• Tags: text-to-speech, mach, on-device

Links

• Model: https://huggingface.co/SyzygyResearch/Mach-Boom
• Files: https://huggingface.co/SyzygyResearch/Mach-Boom/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download SyzygyResearch/Mach-Boom

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