mohammedaly22

/

VoiceTut-TTS

Like 27

Hi Mohammed 👋

Thanks for the GGUF package — one self-contained 1.35 GB file, no Python, and it produced correct Egyptian Arabic on the very first try on a 2014 laptop. I benchmarked it against the PyTorch path on the same machine and got results at both extremes, so I hope the numbers are useful.

Setup — AMD A8 PRO-7600B R7 (4 cores, 2014, no AVX2/FMA, no CUDA), 14.9 GB RAM, Windows 10. Same Egyptian-Arabic text (43 chars), same 6.0 s reference clip + its transcript, --seed 42, --threads 4.

Two observations that surprised me:

• On the CPU path, ggml is 6.6x slower than Python/MKL here (RTF 208.7). My guess: the fast kernels assume AVX2/FMA and q8_0 dequantization on plain SSE is the bottleneck.
• On the same machine the Vulkan build is 1.8x faster than PyTorch at 8 steps and 5.4x faster at 16 steps — even though the driver reports fp16: 0, int dot: 0, matrix cores: none. Really nice work on that backend.

Questions

1. Is AVX2 effectively required for the CPU path to beat the Python/MKL path? If so, would you consider a note in the README (e.g. "CPU-only without AVX2: prefer the Python package")? I'm happy to run more tests (fp32 GGUF, thread counts, 1/4/8 steps) and report back.
2. My CLI wall time (~67 s) looks like it includes ~30 s of model load. Would audiocpp_server with a resident model bring per-request time to ~35 s? I'm about to measure that too.

(And quietly still hoping for the Cairo-TTS weights 🙂 — my CPU-only environment has been ready since Oct 14; I also asked in the other discussion.)

Thanks again — this package is the first thing that made TTS feel fast on this laptop.

— Sami (16), Egypt