Verify TTS pronunciation with Kaldi phoneme decoding: a reproducible fix for polyphonic Chinese characters
dotey · x · 2026-10-22
goldengrape published and verified a reproducible method for checking whether TTS reads polyphonic or rare Chinese characters correctly, replacing unreliable manual listening. Instead of trusting recognized text—where the language model can mask wrong pronunciations—it reads the phoneme sequence from Kaldi's acoustic decoding path and compares it against the target character's correct reading, demonstrated with 攒 (cuán vs zǎn) in 攒竹穴.
Key practical points:
- Runs on CPU only (Ubuntu 24.04 tested), 4 GB RAM, 4 GB disk
- Install Kaldi via conda-forge CPU build rather than compiling from source; LDLIBRARYPATH must include $EL/fst to avoid OpenFST plugin errors
- Uses Vosk's large Chinese model vosk-model-cn-0.22 (a packaged Kaldi chain model); the small model's private lookahead FST format segfaults upstream Kaldi decoders
- Audio is normalized to 16 kHz mono, with the target word appearing multiple times for cross-checking
dotey confirmed the method works, and the article notes you can hand it to an AI agent to complete the setup.
More from coding & agent
- Combining Ghidra, AI Reverse Engineering and WebGPU Sparks an "Absolute Renaissance" — thursdai_pod · 2026-10-22
- As AI writes more code, observability becomes the key to catching agent business-logic failures — mipsytipsy · 2026-10-22
- Reordering Identical C Struct Fields Cuts Size From 24 to 16 Bytes — cneuralnetwork · 2026-10-22
- fal employee shows Claude + fal generating a game in 60 minutes — OdinLovis · 2026-10-22
- Optimizing a TV-screen AI voice assistant UI for a more immersive look — op7418 · 2026-10-22
- GitHub Copilot CLI v1.0.103 ships cached MCP snapshots, sandbox and session-resume fixes — copilot-cli-release-app[bot] · 2026-10-22