License for the model weights (tinytts.onnx, G.pth) and training data
Hi, and thank you for TinyTTS. It is impressively fast for its size.
We are evaluating it as a local voice for Nutq, an open-source (MIT) voice agent that runs in the browser, with ONNX Runtime Web on the CPU. We would like to ship the model with the app, but before we can, we need the licensing of the weights to be explicit. Could you help with a few questions?
License of the weights. Under what license are the files in this repository released (tinytts.onnx, tinytts_fp16.onnx, G.pth and config.json, at revision 6d0cc902621451fcd47462b50ee0aabc6a852e25)? The GitHub repository is Apache-2.0, but that license covers the files in that repository, and these files are only here. If they are Apache-2.0 too, could you add a model card (README.md) with "license: apache-2.0" to this repository, so it is stated where the weights are downloaded?
Training data. config.json lists "train_ljspeech.list". Was the model trained only on the LJ Speech Dataset (public domain)? Or was any other data, or a pretrained or teacher model (for example a MeloTTS checkpoint), used in training or initialization?
Attribution. The symbol table and the language and tone maps in tiny_tts/text/symbols.py match MeloTTS (MIT). Is part of the code derived from MeloTTS? We want to credit both projects correctly.
A small one: the only speaker is labelled "MALE", while LJ Speech is a single female reader. Is the label meaningful, or just a name?
In case it is useful: in ONNX Runtime Web, creating the session for the single tinytts.onnx took 3.7 to 4.8 s in our tests, almost all of it in ONNX Runtime's memory planning for the large graph. Splitting the graph into three parts (encoder and duration predictor, flow, decoder) and pre-optimizing each brought that to 2.2 to 3.7 s, with identical output. We are happy to share the script.
Thanks again!