Sound Tokens — Git skill for Claude Code

sound-tokens

An [Agent Skill](https://agentskills.io) that builds pronunciation **token reels** for dialect and accent coaching.

Give it a folder of short speech clips and a target word for each one, and it produces a single MP3 drill. Each token plays the phrase at normal speed, then the isolated target word slowed with pitch preserved, repeated:

[full phrase @ normal speed]  →  gap  →  [isolated target @ half speed] ×3

Tokens are grouped by lexical set, normalized to a consistent level, and tightly spaced. The reel plus per-token clips are written into the set's own folder. You hear the sound in context, then drill it slowed — which is the thing a student actually needs and the thing that takes an hour to cut by hand.

What it does

• Locates words by forced alignment, not speech recognition. You supply the transcript; pocketsphinx returns per-word timestamps. This is the core design choice — recognition mis-hears strongly accented speech badly, alignment does not, because the word sequence is given rather than guessed.
• Groups by lexical set (KIT, FLEECE, NURSE, NORTH/FORCE …) in one continuous reel.
• Keeps duplicates and clusters them. The same word from different speakers is the comparison the drill exists for. Same-word instances sit adjacent (~0.22 s apart) with a soft breath (~0.45 s) between different-word clusters.
• Pads intelligently at word edges. A voiceless obstruent edge (F S SH TH HH P T K CH) gets extra room so fricatives and stop bursts survive the cut instead of being clipped to a click.
• Slows with atempo, so pitch is preserved — and never double-slows a clip that was already recorded slow.
• Dry-runs first. It prints every boundary it found so you can catch a bad cut before any audio is encoded.
• Different profile for consonant sets. Vowel sets repeat ×3; consonant and feature sets (:r:, :L:, :h:, :t: :d:, clusters, affricates) play ×1 with tighter, lead-free spa