mytts-en-eval: evaluation lists and validation gates

What this repository is: the fixed English test lists that the mytts English models are scored on, and the results of the validation gates that checked the scoring pipeline on real recordings before any model number was published. It holds no model, no training data and no audio. It changes only when the evaluation protocol changes.

• The model: VoiceHub/DACFlow-EN-10k (weights, samples you can listen to, scores)
• Its training data (tokenized): VoiceHub/DACFlow-EN-10k-data
• The evaluation audio these lists point to (eval/en/), in the complete working backup: VoiceHub/DACFlow-EN-10k-backup

Files

file	what
lists/seedtts_en.jsonl	Seed-TTS test-en: 1,088 cases, each a Common Voice prompt (its audio and text) and a target sentence to speak in that voice
lists/seedtts_en_dev.jsonl	the frozen dev half (545 cases): the decision set seed_dev, used to choose between recipes
lists/seedtts_en_test.jsonl	the frozen test half (543 cases): read only at milestones and releases, never for a decision
lists/seedtts_en_split.json	which case is in dev and which in test
lists/lspc_test_clean.jsonl	LibriSpeech-PC test-clean, 4-10 s cross-sentence: 1,127 cases (the F5-TTS protocol)
validation/gates.md, validation/gates.json	the validation gates and their verdict (all passed on 2026-09-29)
validation/<set>_gt/, validation/<set>_resynth/	the scoring of the real recordings (gt) and of their codec resynthesis (resynth): summary.json and one line per case (results.jsonl); <set> = seed_full (Seed-TTS test-en) or lspc (LibriSpeech-PC)

One line of a list (shortened):

{"id": "common_voice_en_103675-common_voice_en_103676", "text": "One by one, the campfires were extinguished, ...", "lang": "en",
 "prompt_audio": "/workspace/data/eval/en/seedtts_testset/en/prompt-wavs/common_voice_en_103675.wav", "prompt_text": "I'm never more aware ...", ...}

prompt_audio is a path on the project's machine: the same file is under eval/en/ in VoiceHub/DACFlow-EN-10k-backup (here: eval/en/seedtts_testset/en/prompt-wavs/common_voice_en_103675.wav).

How a model is scored

The model reads text in the voice of prompt_audio with the fixed EN-E1 settings (the length of the output comes from the band rule, never from the reference recording). Then:

• WER / CER: openai/whisper-large-v3 (greedy, English) transcribes the output, both sides are normalized with the frozen en-v2 rules (fillers, apostrophes, spelling variants), errors are counted over the whole list (corpus WER);
• SIM-o: WavLM-large + ECAPA-TDNN cosine between the output and the prompt (the Seed-TTS-eval / F5-TTS scorer);
• UTMOS: UTMOS22 strong, a predicted naturalness score.

Two models are compared with a paired bootstrap over the same cases (95 % confidence intervals). The full protocol: docs/EVAL_EN.md in kadirnar/dacvae-next (branch roadmap/en-echo; private at the moment).

What the validation gates checked

Before any model was scored, the same pipeline scored the real recordings of these lists. The numbers had to match the published ones, so that a model's numbers mean what they say (validation/gates.md):

gate	measured	target
Seed-TTS test-en, real audio, WER (seed-tts-eval protocol)	2.14 %	2.13 % +- 0.15	PASS
LibriSpeech-PC test-clean, real audio, WER (F5-TTS protocol)	2.47 %	2.40 % +- 0.20	PASS
Seed-TTS test-en, real audio, SIM-o to the prompt	0.734	0.734 +- 0.005	PASS
LibriSpeech-PC, real audio, SIM-o to the prompt	0.695	0.695 +- 0.005	PASS
Seed-TTS test-en, real audio, UTMOS	3.523	3.53 +- 0.05	PASS
LibriSpeech-PC, real audio, UTMOS	4.098	4.10 +- 0.05	PASS

The resynth runs pass the same audio through the codec only (encode + decode): their WER stays within a tenth of a point of the real audio's (en-v2 corpus WER), so the codec itself costs no intelligibility.

Not here

• The echo lists (sentences and prompts of the 236 held-out voices of SynDataLab-EN/echo-clones-4m-en): with their audio in VoiceHub/DACFlow-EN-10k-backup, eval/en/echo_heldout/ and eval/en/echo_heldout_audio/.
• The AMI lists (ami_ihm.jsonl, ami_sdm.jsonl) and their audio: VoiceHub/DACFlow-EN-10k-backup, eval/en/.

Licence

The lists are derived from Seed-TTS-eval (zhaochenyang20/seed-tts-eval, CC BY 4.0; its prompts come from Common Voice) and LibriSpeech / LibriSpeech-PC (openslr/librispeech_asr, CC BY 4.0), and keep that licence and attribution. The validation results are the project's own.