yomiscript-ja-pitch

The pitch model of @yomiscript/ja-accent (pitchYomite, decisions D33 and D35 of YomiScript): a two-layer bidirectional LSTM (2.12 million parameters) that reads the morae of a Japanese sentence with a dictionary's accent nuclei, and predicts three pitch values per mora. It runs in plain TypeScript, in Node and in the browser.

qwen-2/ holds pitch.json.gz (vocabularies, tensor shapes, the linear pitch model it adds to) and pitch.bin.gz (the weights, little-endian float32). qwen-1/ is the previous version (1.78 million parameters, 3,312 sentences), kept for the package versions that load it.

Trained on the pitch of Qwen3-TTS (1.7B Base, a cloned voice) reading 1,144 even-numbered JSUT BASIC5000 sentences and 4,763 Japanese Wikipedia sentences; 368 more chose the epoch. Its vocabulary holds words from those texts, so it is under CC BY-SA 4.0, as they are. receptron/yomiscript-ja-qwen-speech holds the speech of the sentences of qwen-1, not the 2,595 Wikipedia sentences added for qwen-2.

Distance to Qwen's pitch on 1,214 sentences it did not see (semitones, each sentence centred): the dictionary's nuclei 1.871, qwen-1 1.592, qwen-2 1.543, nuclei read from Qwen's own pitch of each sentence 1.494. Blind, 12 sentences and 2 repeats, one listener: 13 the same as qwen-1, 1 better, none worse. Rebuilt by eval/qwen_prosody/more_train.py.