Breeze TTS 2 fine-tuning by Instavar
Collection
Source, LoRA and full-SFT releases from the Instavar Breeze TTS 2 Singapore English adaptation study. • 3 items • Updated
You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This derivative is available only for research and non-commercial use under the included BreezeBlue agreement. Gating records an acknowledgement but does not replace the agreement or supply voice and recording rights.
Log in or Sign Up to review the conditions and access this model content.
Instavar SG Narration Full SFT
Derived from Breeze TTS 2 by BreezeBlue and licensed for research and non-commercial use only.
This is an independent research checkpoint trained by Instavar on a consented, single-speaker subset of Singapore's National Speech Corpus. It is not an official BreezeBlue release and is not endorsed by BreezeBlue or RESONIA, INC.
Artifact
- Base:
BreezeBlue/Breeze-TTS-2 - Pinned base revision:
799624c0b4a1daa8db6d28bbd9850043c0270734 - Selected checkpoint: step 750 of 1,000
- Updated synthesis parameters: 2,387,151,872
- Frozen during training: text encoder and audio codec
- Optimizer: FP32-master SGD
This repository contains inference roles only. Optimizer, scheduler, random state, trainer state, source recordings, caches, and private receipts are not included.
Use
Use the companion source toolkit:
python infer.py /models/sg-narration-full-sft \
--text "The train arrives in five minutes." \
--output output.wav
Source and documentation: instavar/breeze-tts2-finetuning
Experiment report: Breeze TTS 2 LoRA and Full Fine-Tuning for Singapore English
Related model: Instavar SG Narration LoRA R8
Collection: Breeze TTS 2 fine-tuning by Instavar
Evaluation
In a matched reference-free comparison, this checkpoint reached mean ECAPA speaker similarity 0.6973 and WER 0.0467. The LoRA adapter reached 0.6810 and the same WER. The paired ECAPA interval crossed zero, so the objective study did not establish a winner.
One listener preferred LoRA for cadence and long-form listening, with three short-prompt ties. Both releases mispronounced a tested local word. These results do not establish faithful identity cloning, general Singaporean-accent coverage, production fitness, or broad listener preference.
Licence and responsible use
The included BreezeBlue agreement permits research and non-commercial use only. Do not use this artifact for a product, paid service, client work, advertising, revenue generation, production deployment, impersonation, deception, or voice cloning without legally sufficient consent and recording rights.
Review LICENSE, NOTICE, PROVENANCE.json, and SHA256SUMS before use.
- Downloads last month
- -
Model tree for instavar/sg-narration-full-sft
Base model
BreezeBlue/Breeze-TTS-2Collection including instavar/sg-narration-full-sft
Evaluation results
- Word error rate on FEMALE_01 matched reference-free promptsself-reported0.047
- Mean ECAPA speaker similarity on FEMALE_01 matched reference-free promptsself-reported0.697