espnet3.systems.f5tts.src.pinyin.text_to_pinyin_ids

source

espnet3.systems.f5tts.src.pinyin.text_to_pinyin_ids(text: str, vocab_char_map: Dict[str, int]) → ndarray

F5’s exact mapping: tokenize then map each token, unknown -> 0.

• Parameters:
  • text – One raw transcript.
  • vocab_char_map – Mapping from load_vocab_char_map().
• Returns: Token ids as an int64 array.

Example

>>> text_to_pinyin_ids("abz", {" ": 0, "a": 1, "b": 2})
array([1, 2, 0])

NOTE

F5’s vocabulary has no <unk>, so an unknown token maps to index 0, which is the literal space. That is upstream’s behaviour and is kept deliberately for checkpoint compatibility, but it means out-of-vocab text degrades silently rather than raising.