Phoneme alignment support

[JarbasAl]

Goal

Add phoneme-level timing alignment so consumers can map audio samples back to individual phonemes.

What is phoneme alignment

Per-phoneme start/end sample indices extracted from the model's duration predictor output. Enables use-cases like viseme lip-sync, karaoke highlighting, and pronunciation analysis.

Upstream

Ported from piper (OHF-Voice/piper1-gpl commit eb6be6b). The alignment logic does not use espeak-ng and is covered by Apache 2.

Design

• Second ONNX output tensor: per-frame phoneme IDs
• hop_length (default 256) converts frames → samples
• PhonemeAlignment dataclass: phoneme + num_samples
• AudioChunk extended with phoneme_alignments, phoneme_ids, phoneme_id_samples
• TTSVoice.synthesize(include_alignments=True) returns populated chunks when the model supports it
• phoonnx_train export-onnx --add-phoneme-alignment embeds the extra output in the exported graph

PR

#125 — feat/phoneme-alignment

Deliverables

• PhonemeAlignment dataclass
• AudioChunk alignment fields
• synthesize / phoneme_ids_to_audio alignment path
• tests/test_alignment.py (24 hermetic unit tests)
• docs/alignment.md (usage guide with viseme and karaoke examples)

Acceptance criteria

• feat: phoneme alignment support #125 merged
• ONNX export with --add-phoneme-alignment produces valid two-output graphs
• synthesize(include_alignments=True) returns PhonemeAlignment lists when model supports it
• synthesize(include_alignments=True) returns None fields gracefully when model does not support it