feat(wake-word): support Chinese wake phrases via sherpa wenetspeech pinyin model

[ruhipiano]

Summary

The sherpa wake-word engine hardcoded BPE tokenization (text2token + bpe.model), which only covers Latin-script phrases on the default English GigaSpeech model. The wenetspeech KWS model — sherpa-onnx's Chinese counterpart — is pinyin-modeled (声母/韵母 phonemes) and ships no bpe.model, so pointing wake_word.sherpa.model_dir at it crashed at listener init.

Changes

• Model-family detection: _SherpaKwsEngine now checks for bpe.model — BPE models keep the existing text2token path; pinyin models (no bpe.model) romanize phrases to phonemes.
• Pinyin tokenizer (_phrase_to_pinyin_phonemes): splits pypinyin's toned syllables (Style.TONE) after the initial — nǐ → n ǐ, wén → w én, huì → h uì — which also fixes pypinyin's mis-timed bare finals (úen, ueì), with a ü→u fallback for finals the vocab spells without umlaut (üé → ué).
• Graceful degradation: phrases a model cannot represent (e.g. Latin on the pinyin model) are skipped with a warning instead of failing the whole listener — so mixed-language profile routing degrades safely rather than crashing.

Verification

• Unit tests: 31 pass (5 new pinyin-tokenization tests, offline-safe via tmp_path + pytest.importorskip).
• E2E against the real wenetspeech model + its bundled test wavs: all 8 official keywords recognized (文森特卡索/周望军/朱丽楠/蒋友伯/女儿/法国/见面会/落实) with correct phoneme output matching the official keywords.txt byte-for-byte.
• TTS-rendered 你好妮妮 fires the listener with correct profile routing (last_match == ('你好妮妮', 'nina')).

Docs

New "Chinese wake phrases (pinyin model)" section under the sherpa engine docs with model download link and config example.

Notes

• pypinyin was already a wake.sherpa lazy-dep (sherpa_onnx's text2token imports it unconditionally, so no new dependency.
• The wenetspeech model (~13 MB) is auto-downloadable from sherpa-onnx releases; users point at the unpacked directory.
  EOF
  )

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

PR: feat(wake-word): support Chinese wake phrases via sherpa wenetspeech pinyin model

1. zip(initials, syllables) in _phrase_to_pinyin_phonemes (wake_word.py:119) silently truncates if pypinyin returns different-length lists for a mixed phrase (Chinese + Latin/digits), which would misalign phoneme output for the remainder of the phrase. Consider guarding len(initials) == len(syllables) and skipping the phrase (or the unaligned tail) instead of zipping blindly.
2. _sherpa_token_vocab re-reads and re-parses tokens.txt on every phrase in __init__ (per enrolled profile + config phrase), i.e. O(phrases × vocab file) I/O on each listener start. Reading the vocab once in the constructor and passing it to the tokenizer would be simpler and cheaper.
3. if not phrase.isascii(): return [] (wake_word.py:155) drops accented-Latin phrases ("héllo", "café") even though the BPE vocab may contain them. Consider allowing Latin-1/Unicode letters (or checking against the vocab) instead of ASCII-only, since the docstring says BPE covers "Latin-script phrases".
4. The ü→u fallback only fires when rest not in vocab, so it can't corrupt a vocab that spells üè correctly — good. But a vocab spelling neither üé nor ué leaves the final silently dropped with no warning; a debug log for skipped finals would help users diagnose unexpected phrase degradation.