VoiceHub/mytts-en-base

Results: English TTS (mytts-en)

English zero-shot TTS (flow-matching DiT on DACVAE latents) of kadirnar/dacvae-next (branch roadmap/en-echo, protocol EN-E1). This board is written by scripts/promote.py (EN-46) from docs/RESULTS_EN.json; every candidate model goes through one gate, and only a promoted model is released to VoiceHub/mytts-en.

Promotion gate (PLAN §1 decision rule). A candidate is compared with the current best on the two decision sets of the English protocol, Whisper-large-v3 greedy with en-v2 corpus WER: since protocol version 2 (#205: each prompt's own bandwidth, the sim3 column) echo-dev v2 (500 held-out DEV-voice sentences x prompt draws d0 + d1, text-disjoint from training) and seed-dev (the frozen dev half of Seed-TTS test-en x sampling seeds 0 + 1); version-1 entries used echo-dev (130 held-out-voice sentences x d0 + d1). It is promoted when its paired corpus-WER gain (95 % bootstrap CI over sentences / prompt-connected components) excludes 0 on at least one set, it is significantly worse on neither, SIM-o drops by at most 0.01, UTMOS by no more than the seed spread and (version 2) sim3 does not drop significantly while SIM-o holds. Version 2 also has a real-voice guard with a veto: a candidate whose share of seed-dev outputs below SIM-o 0.2 is significantly above 5 % (its 95 % CI's lower bound above it) is rejected, also the first model (since 2026-10-16: a lead decision under the owner's delegation, based on the re-baselining of the history under version 2). A promoted model then gets the full EN-E1 table on the held-back sets (read only here and at releases) and must not regress on any of them; then it is released (safetensors, model card with data provenance and licence notes, AudioSeal output watermark checked on its own samples).

Current best: grow-q-u300

Release	not released
Source	run grow_q_u300, step 300 (VoiceHub/mytts-en-base@93b162c3e3cb920712ac18a59c10ca8c1855d277:grow_q/checkpoints/model_step_00000300.pt)
Change	first English board entry (protocol v2): the shipped model, EN-45 GROW quality round update 300 on the EN-34 100k EMA (VoiceHub/DACFlow-EN-10k)
Promoted	2026-10-17 01:05 UTC over (none: first model): promote (first model: no current best)
Code	34abc48

EN-E1 table (held-back sets)

Read only at promotions and releases, never for a decision. Unit of the CIs: sentences (echo), prompt-connected components (Seed, LS-PC, echo pairs). Published protocol: seed-tts-eval on Seed, F5 on LibriSpeech-PC (per-utterance means, as published). GT floor: the ground truth through the same scorer.

set	utt.	WER (en-v2)	95 % CI	published protocol	WER > 50 %	SIM-o	SIM-2	UTMOS	DNSMOS	judge 2	<3s	3-5s	5-10s	>10s	GT floor
echo-test	390	0.64	0.36-0.99	-	0.0	0.788	-	3.90	3.17	0.23	1.17	0.97	0.44	0.50	-
seed-test	1629	1.64	1.31-2.00	2.11 (seed-tts-eval)	0.0	0.546	0.612	3.86	3.11	1.22	0.84	1.13	2.65	-	1.51
Seed test-en (1,088)	1088	1.46	1.22-1.70	1.97 (seed-tts-eval)	0.0	0.543	0.611	3.85	3.10	1.22	0.88	1.14	2.12	-	1.55
LibriSpeech-PC test-clean	1127	1.98	1.62-2.33	2.40 (F5)	0.1	0.522	0.572	4.06	3.22	1.42	-	1.60	2.06	0.00	1.87
echo pairs	139	1.33	0.83-1.91	-	0.0	0.776	-	3.88	3.22	0.97	2.56	1.35	1.13	1.27	1.52
AMI wild: headset (IHM)	498	4.47	3.69-5.25	-	0.0	0.406	-	3.84	3.08	2.78	-	4.26	4.56	13.79	9.28
AMI wild: far field (SDM)	498	4.11	3.23-4.97	-	0.0	0.024	-	3.36	2.80	2.63	-	3.77	4.29	13.79	13.08

Targets

target (PLAN §1, proposed)	current best	M4 target	M5 target	status
echo-test, en-v2 corpus WER (3 draws)	0.64 %	<= 2.50 %	<= 2.00 %	M5 met
Seed test-en full, seed-tts-eval WER	1.97 %	<= 2.60 %	<= 2.13 %	M5 met
LibriSpeech-PC test-clean, F5 WER	2.40 %	<= 3.00 %	<= 2.50 %	M5 met
UTMOS on Seed test-en	3.85	>= 3.50	>= the M4 value	M4 met

Promotion history

#	date	candidate	vs	echo-dev dWER pp [CI]	seed-dev dWER pp [CI]	status	verdict	release
1	2026-10-17 01:05 UTC	grow-q-u300 (grow_q_u300 @ 300)	-	- (v2)	-	promoted (current)	promote (first model: no current best)	-
2	2026-10-17 01:06 UTC	base-190k (en-base-s1 @ 190000)	grow-q-u300	-0.06 [-0.19, +0.05] n.s. (v2)	+0.07 [-0.14, +0.26] n.s.	rejected	reject: no significant gain; real-voice GUARD FAILED on seed_dev: 34.2 % [29.9, 38.6] of the outputs below SIM-o 0.2, significantly above 5.0 % (veto)	-

base-190k: rejected

Change: deliberately worse candidate (#70 acceptance): EN-34 en-base-s1 190k, the last pretraining export (#205: 34.2 % of seed_dev outputs below SIM-o 0.2)

reject: no significant gain; real-voice GUARD FAILED on seed_dev: 34.2 % [29.9, 38.6] of the outputs below SIM-o 0.2, significantly above 5.0 % (veto)

set	unit	WER best -> candidate (%)	dWER pp [95 % CI]	dSIM-o	dUTMOS	UTMOS rule
echo_dev_v2	sentence	0.50 -> 0.44	-0.06 [-0.19, +0.05] n.s.	+0.003	+0.105	no significant paired drop (eval.md §8.4; no seed spread recorded)
seed_dev	component	1.42 -> 1.49	+0.07 [-0.14, +0.26] n.s.	-0.236	-0.397	no significant paired drop (eval.md §8.4; no seed spread recorded)

Real-voice guard (seed_dev): 34.2 % [29.9, 38.6] of the candidate's outputs below SIM-o 0.2 (threshold 5.0 %): REAL-VOICE GUARD FAILED (veto: the candidate is rejected).

sim3 (WeSpeaker) candidate - best: echo_dev_v2 +0.0015 [-0.0006, +0.0037]; seed_dev -0.2366 [-0.2532, -0.2199]

• real-voice GUARD FAILED on seed_dev: 34.2 % [29.9, 38.6] of the outputs below SIM-o 0.2, significantly above 5.0 % (veto)

grow-q-u300: promoted

Change: first English board entry (protocol v2): the shipped model, EN-45 GROW quality round update 300 on the EN-34 100k EMA (VoiceHub/DACFlow-EN-10k)

promote (first model: no current best)

set	unit	WER best -> candidate (%)	dWER pp [95 % CI]	dSIM-o	dUTMOS	UTMOS rule
echo_dev_v2	sentence	- -> 0.50	-	-	-	-
seed_dev	component	- -> 1.42	-	-	-	-

Real-voice guard (seed_dev): 1.7 % [0.8, 2.8] of the candidate's outputs below SIM-o 0.2 (threshold 5.0 %): ok.

Held-back check: first model: nothing to compare