license: mit
language:
- zh
tags:
- text-to-speech
- evaluation
- webfm
webfm-eval
Fixed evaluation set and harness for the WebFM TTS work. Small on purpose: it exists so two model tiers can be compared on the same sentences in the same run, not to be trained on.
Contents
eval/sentences.json 40 sentences, 7 buckets (short / medium / long /
number / latin / polyphone / narration)
eval/score.py raw + normalised CER, prefix CER, per-bucket and
per-voice aggregation, delta against a baseline
eval/t2s.json traditional -> simplified table, used only when
opencc is unavailable
config/ming_voices.json the eight WebFM voices and their prompt transcripts
refs/wav/v01..v08.wav prompt clips, reconstructed from the shipped F5
reference mels
kaggle/bake_voices.py prompt_latent + CAM++ speaker_emb for each voice
kaggle/ming_eval.py teacher x student matrix, resumable per cell
Why the scoring is fussy
Whisper usually transcribes Chinese in traditional characters. Comparing that
against simplified references and counting character differences adds tens of
points of fake error. Every score here is reported twice: cer_raw (as
produced) and cer_norm (after traditional->simplified, punctuation stripped).
The gap is measurement, not model quality.
cer_prefix scores only the first four characters. A garbled opening syllable
barely moves overall CER but ruins a radio line, and it is the failure mode that
weight-only int4 quantisation produced on the 3-step student.
Usage
python3 eval/score.py selftest
python3 eval/score.py aggregate --results results.jsonl --baseline teacher --by-bucket
results.jsonl carries one row per generated utterance:
{id, model, voice, text, hyp_text, spk_cos?, mel_l1?, dur_ratio?}.
Licence
MIT for the code. The prompt clips derive from a purchased SFW corpus; check the source terms before commercial use.