A Tiny Speech Judge That Agrees With My Ears
272 blind A/B verdicts, a fixed model of the ear and a few hundred parameters: a judge that predicts my listening better than CDPAM or UTMOS, why anchoring it to human speech …
Category archive
Machine Learning
Field notes collected under Machine Learning.
10
entries
→
Compressing Kokoro-82M: Most of the Decoder Can Go, Its Last Layer Cannot
Pruning and distillation shrink Kokoro-82M to 59.3M parameters with no audible difference in a blind test, and to 45.2M at 1.74x decoder speed. Why one trick made phase …
Teaching Qwen3-TTS to speak Swahili
Qwen3-TTS knows ten languages and Swahili is not one of them. This is how a small adapter taught it, what the tokenizer had to be told about sounds like mb and ng', and why …
The architecture of Qwen3-TTS-12Hz: two summed streams, 16 codes per 80 millisecond frame, 2.10 B parameters
A read of the Qwen3-TTS-12Hz checkpoint from its safetensors headers rather than its model card, covering the talker, the 15 step depth predictor, and the RVQ decoder that …
Running Qwen3.8 9B on a 16 GB M4: 36.4 tok/s Decode and a 57x Prefill Cut
Which quantisation fits 16 GB, why prefill is already at a hardware limit, and the two defects that stop prompt caching from working. Prefill on a 4,312 token prompt falls …
Taking the 2B Qwen3-TTS from 0.876 RTF to 0.253 RTF on Apple Silicon M4
A day of profiling a two-billion-parameter speech model on an M4 MacBook Pro, told as six wrong predictions. The largest win was nine lines of sorting, and one of my two …
Tuning CosyVoice3 in Rust: Where the Speed Actually Was
The same speech model made fast a second time, in Rust on Metal, from RTF 2.96 to 0.697. Scheduling beat kernels, and two of the measurements I steered by were wrong.
How I Made CosyVoice3 2.2× Faster on Apple Silicon
How I bounded PyTorch MPS memory growth, cut CosyVoice3 latency below real time, and moved only the dispatch-bound LLM stage to MLX.
From Noisy Kenyan ID OCR to JSON with MLX and Candle
Fine-tune a 50M-parameter model on synthetic Kenyan ID OCR, measure field-level accuracy, and package the fused F16 weights for Candle inference in Rust.
From PP-OCRv6 Text to Kenyan ID JSON
A measured walkthrough of the modes, schema, synthetic fixture, and small understanding model I put around PP-OCRv6 to extract Kenyan ID fields locally.
No entries are filed here yet.
Browse all notes →