mytts-en-base

mytts-en-base: internal training state of the main run

Internal storage, not the model page. This repository holds the working files of the English main training run (en-base-s1: the 206 M base model, about 10,000 h of SynDataLab-EN/echo-clones-4m-en, 200k steps) and of the runs forked from it: resumable training states, EMA weights every 10k steps and their evaluations. They are here so that a run can continue after the machine is lost. Released models are not published here.

• The model (released weights, samples you can listen to, how to use it): VoiceHub/DACFlow-EN-10k
• Its training data (tokenized): VoiceHub/DACFlow-EN-10k-data

What is here

The runs write here while they train; each run has its own folder (its train.hub_prefix).

How the main run is set up

• Model: base preset, 206 M parameters: a flow-matching DiT that generates the codec latents of Aratako/Semantic-DACVAE-Japanese (25 Hz x 128).
• Data: the catalog echo_en_v1 of VoiceHub/DACFlow-EN-10k-data: every clip of echo-clones-4m-en that passes the quality filters, the 236 held-out and 177 echo-unseen voices left out. Nothing else: the run trains only on the provided data.
• Recipe: 200k steps; noise schedule t_mean -0.8 / t_std 0.8 with the EN-29 prompting (en29-fixed-pc06), both chosen in the screening runs of VoiceHub/mytts-en-ablations; REPA alignment to w2v-BERT 2.0 layer 16. Configuration: configs/train/en_full.yaml in kadirnar/dacvae-next (branch roadmap/en-echo; private at the moment).

Licence

apache-2.0 (owner decision, 2026-09-29). Trained only on SynDataLab-EN/echo-clones-4m-en (apache-2.0 on its card), synthetic speech generated with EchoTTS (jordand/echo-tts-base, cc-by-nc-sa-4.0 on its card). The provenance notes are on the model page.