đ¸ Takane
â ī¸ Warning: This demo doesn't have load balancing or parallel query handling and it is under heavy load. Please try again later if it didn't work. Sorry!
ããŽããĸã¯äģããŠããŖãã¯ãå¤ããĻãéčĻãĢå ¨é¨å¯žåŋã§ããĒããããããžãããããžãåäŊããĒããŖãå ´åã¯ããã˛åžãģãŠããä¸åēĻãčŠĻããã ããã
WARNING! Longform Generation is not optimized for this mode
Duration Control Parameters
don't touch - For legacy compatibility
Sample Text and Audio Prompts
Click on any example below to load it into the Speech Generation tab
Note: Temperature and chained longform settings are loaded from the examples file
About Takane
Takane is a frontier Japanese-only speech synthesis network that was trained on tens of thousands of high quality data to autoregressively generate highly compressed audio codes. This network is powered by Kanadec, the world's only 44.1 kHz - 25 frame rate speech tokenizer which utilizes semantic and acoustic distillation to generate audio tokens as fast as possible.
There are two checkpoints in this demo, one of them utilizes a custom version of Rope to manipulate duration which is seldom seen in autoregressive settings. Please treat it as a proof of concept as its outputs are not very reliable. I'll include it to show that it can work to some levels and can be expanded upon. Both checkpoints have been fine-tuned on a subset of the dataset with only speaker tags. This will allow us to generate high quality samples without relying on audio prompts or dealing with random speaker attributes, but at the cost of tanking the zero-shot faithfulness of the model.
Takane also comes with an Anti-Hallucination Algorithm (AHA) that generates a few candidates in parallel and automatically returns the best one at the cost of introducing a small overhead. If you need the fastest response time possible, feel free to enable the Turbo mode. It will disable AHA and tweak the parameters internally to produce samples as fast as 2-3 seconds (though due to an influx of users coming in, you probably will be qeued and have to wait!)
There's no plan to release this model for now.
If you're not using an audio prompt or a speaker tag, or even if you do, you find the later sentences to be too different, then in that case you may want to enable the Chained mode, which will sequentially condition each output to ensure speaker consistency.
Summary of Technical Properties:
- Encoder-Decoder fully autoregressive Transformer
- Powered by Kanadec (44.1 kHz - 25 codes per second)
- 500M parameters
- Tens of thousands of anime-esque data, everyday regular Japanese is not supported
- Experimental support for duration-controllable synthesis
đ¸ Takane - Advanced Japanese Text-to-Speech System