mbfa-japanese (p4mbfa · japanese)

A context-free CNN phoneme encoder and forced aligner (p4mbfa, the successor of CUPE / p3cupe), for the japanese language group of standard_g2p. Each 5 ms frame is classified from at most 120 ms of audio (38.9 ms receptive field), so the model cannot learn any language's phonotactics. Phone boundaries come from a segmental Viterbi decoder over the known phone sequence, refined to sub-frame precision at the crossing of neighbouring posteriors.

head	labels	size
ph	local tokens of japanese (incl. <blank> SIL noise <unk>)	28
phg	gold phoneme groups, shared by every language group	15
tone	tone values, shared by every language group; present but untrained (no language in this group has a tone layer)	22

Languages (the FLEURS languages it was trained on): ja. standard_g2p maps every member of the group onto the same tokens, so other members may align too, untested.

Training

Data: FLEURS japanese, all 7.0 h, training noise_level 0.02.
Checkpoint: experiment mj01a, epoch 8, selected on FLEURS val_loss and frame accuracy (epoch 8 of 30).
Trunk and phg_head from latin ma02a (final epoch), ph_head and tone_head trained fresh.

metric	value
val_loss	2.3491
val_frame_acc	0.5036
val_frame_acc_groups	0.5381

val_* are on held-out FLEURS clips but are scored against the model's own alignments (there are no boundary labels outside English), so they measure self-consistency, not boundary accuracy.

Labels are standard_g2p dictionary pronunciations (gold inventory 9438371ed6dd), not phonetic transcriptions of what was said.

Usage

The code is in https://github.com/tabahi/bfa_models (p4mbfa/); clone it and run from its root.

from p4mbfa.inference import MbfaAligner

aligner = MbfaAligner.from_pretrained("Tabahi/mbfa-japanese")
wav = aligner.load_audio("clip.wav")                     # mono, 16000 Hz
phones = ["SIL", "a", "s", "i", "e", "SIL"]   # this group's tokens (config.json labels.tokens)
for seg in aligner.align(wav, phones):
    print(seg["token"], seg["start_ms"], seg["end_ms"])

aligner.encode(wav) returns the per-frame log posteriors of all three heads. Text -> tokens is standard_g2p's job (goldG2P.phonemize_sentence then lang_group_inventory.to_local); config.json lists the token strings. from_pretrained downloads only config.json and model.safetensors.

Fine-tuning

ckpt/japanese_fleurs7h_mj01a_e8_val_loss=2.349.ckpt is the training checkpoint (PyTorch Lightning, weights only, a pickle: load it only if you trust this repo). Download it, then point a p4mbfa yaml at it to continue training, or to train a new language group on its trunk and shared phg head:

hf download Tabahi/mbfa-japanese "ckpt/japanese_fleurs7h_mj01a_e8_val_loss=2.349.ckpt" --local-dir tmp/hf/mbfa-japanese

ckpt_path: "tmp/hf/mbfa-japanese/ckpt/japanese_fleurs7h_mj01a_e8_val_loss=2.349.ckpt"
reset_fine_heads: true      # new lang_group: rebuild ph_head / tone_head; false = same group

License: AGPL-3.0.