Mach Boom is a version of Gradium's Phonon 31m English text-to-speech model, compressed to 3.5 bits per weight, with a model file size of 18.2 MB. The authors measured approximately 8.3–8.4 times real-time speed on a single core of a 4-core Xeon virtual machine, with peak memory usage of 93 MB and first-segment audio latency in streaming mode of 61–78 milliseconds. On a set of 250 difficult English sentences containing numerous numbers, dates, amounts, and IDs, the authors report a word error rate of 5.18%, which is lower than the 8.61% rate of the original fp32 model. It comes with 4 built-in voices, and voice files from phonon-voices can also be used.
Why it matters Its small size and lack of GPU requirements make it suitable for embedding in local applications or edge devices to read out difficult text such as dates, amounts, and IDs.
Hugging Face likes 3 · Hugging Face trending score 3 · Hugging Face downloads 0
License: Commercial use allowed · ComfyUI: Not available · VRAM: Estimated ≈1 GB
Reasons
- The Apache-2.0 license permits commercial use; the testing methodology, transcription model, and comparison baselines are all documented in the model card.
Usage
- After running
hf download SyzygyResearch/Mach-Boom, install dependencies withpip install numpy tokenizers, then runpython tts.py "text" -o out.wav --voice Zoey. The first run will compile kernels for the local CPU.
Risks
- Supports only Linux x86-64 and English; requires gcc or clang.
- Evaluations were conducted by the authors, and the difficult sentence set comes from Gradium's own dataset.