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 …
Field note index
All notes
Field notes on software, infrastructure, AI, electrical systems, and the work of building and repairing them.
55
entries
→
Synchronizing a Google Pixel Watch With an iPhone
iOS gives Wear OS no companion protocol and no Google Mobile Services, but it leaves Bluetooth LE and ANCS open. A Pixel Watch 4 can push steps, distance, heart rate, resting …
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 …
Jaeger 2.20 and Prometheus: A Two Container Tracing Setup With RED Metrics
A working compose file for Jaeger v2 with on-disk storage, span-derived rate, error and duration metrics in the Monitor tab, and per-environment labelling. Complete configs, …
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 …
Cutting a Cold cargo check From 23.77s to 0.47s on an M4
A cold cargo check of rust-analyzer goes from 23.77s to 0.47s. Three changes across cargo and rustc, what each one bought, and how two of them ship as a separate tool that …
Driving a Phone Through Its Accessibility Tree
Every Android app already publishes a live, structured description of its screen, for screen readers. That description is an API, and it turns a phone into something an agent …
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.
Narrating Book-Length Text on One Machine
A pipeline for turning a book-length document into narrated audio with word-level timings on one machine: three engines behind one trait, the text rules that decide quality, …
How to Receive Email at Home Through a VPS Gateway
A complete Postfix, Dovecot, Rspamd, HAProxy, and reverse-SSH setup for receiving Internet email on a home server without router access.
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.
