phonon-31m-en-small

gradium/phonon-31m-en in 43.2 MB instead of 64.5 MB, with the same voice. Same network, tokenizer, voice and runtime; only the bits per weight changed.

Blind test

Same sentence, same seed. One clip is the official 64.5 MB model, the other is this build.

A

B

Which one is 43.2 MB?

A is this build. B is the official model.

What changed

Part	Official	This build
Transformer inside the frame-by-frame loop	Q8_0	Q8_0
Flow head residual blocks	BF16	Q5_K
Flow head time, final and conditioning layers	BF16	Q8_0
Text and voice conditioning tables	BF16	Q8_0
Mimi decoder transformer	BF16	Q6_K
Mimi convolutions	BF16	F16
Tensors under 16k elements	BF16	F32

The loop transformer keeps 8 bits because its output is fed back in as the next frame's input, so errors compound; at 6 bits the voice drifts. The text and voice tables need 8 bits too. The decoder runs once per frame and takes 6 bits. The runtime cannot block-quantize the convolutions (their last dimension is the kernel size), so they stay 16-bit, as F16, which keeps 3 more mantissa bits than the official BF16.

Measured

3 seeds × 24 sentences per model. Distance is the DTW log-mel gap between a model's takes and takes of the official model at 3 other seeds (9 pairs, lower is closer). Word error is from an ASR model.

	size	distance	word error
Original weights (f32)	102.8 MB	1.972	4.41%
Official model.q8.gguf	64.5 MB	1.968	4.29%
This build	43.2 MB	1.988	3.80%

Smaller builds drifted on the same test: the loop transformer at 6 bits (38.6 MB) +0.048 with all 9 pairs higher, its layers 0-2 at 4 bits (36.2 MB) +0.144, and the text and voice tables at 4 bits about +0.03.

Usage

Same runtime as the parent. Point it at this repo:

uvx ptts -m "hf://edge37/phonon-31m-en-small" --lang en --quant q8 "Hello from Phonon." -o speech.wav

import ptts

tts = ptts.TTS("hf://edge37/phonon-31m-en-small", lang="en", quant="q8")
tts.save("speech.wav", "Hello from Phonon.")

This repo ships the GGUF runtime build only; it has no model.safetensors.

How it was made

shrink_phonon.py streams the parent's original weights from the Hub and writes this GGUF with llama.cpp-compatible Q8_0 / Q6_K / Q5_K blocks. Notes for upstream: gradium-ai/xn-ptts#328.

License

Apache 2.0, inherited from gradium/phonon-31m-en.