SoundsRight (SN)

| Website & Leaderboard | Bittensor |

Bittensor's Speech Enhancement Subnet

If you are unfamiliar with how Bittensor works, please check out this primer first!

The SoundsRight subnet incentivizes the research and development of speech enhancement models through daily fine-tuning competitions, powered by the decentralized Bittensor ecosystem.

Miners in the subnet will upload their fine-tuned models to HuggingFace, and the subnet's validators are in charge of downloading the models, benchmarking their performance and rewarding miners accordingly.

Each competition is winner-takes-all.

Fine-Tuning Competitions

The table below outlines competitions currently being held by the subnet. Competitions are distinguished by the sample rate of the testing data, the task and the metric used for benchmarking.

Sample Rate	Task	Benchmarking Metric	% of Total Miner Incentives
16 kHz	Denoising	PESQ	15
16 kHz	Denoising	ESTOI	12.5
16 kHz	Denoising	SI-SDR	7.5
16 kHz	Denoising	SI-SAR	7.5
16 kHz	Denoising	SI-SIR	7.5
16 kHz	Dereverberation	PESQ	15
16 kHz	Dereverberation	ESTOI	12.5
16 kHz	Dereverberation	SI-SDR	7.5
16 kHz	Dereverberation	SI-SAR	7.5
16 kHz	Dereverberation	SI-SIR	7.5

For more details about sample rates, tasks and metrics, please reference the competition docs.

Getting Started

To get started with mining or validating in the subnet, please reference the following documentation:

Subnet Documentation

| Subnet Architecture | Competitions | Roadmap | Citations |

Miner Documentation

| Running a Miner | Generating a Fine-Tuning Dataset | Model Formatting |

Validator Documentation

| Running a Validator | Validator Architecture |

Running on Testnet

Recently the Bittensor testnet has been updated with dynamic TAO, and a few changes have been made to the functionality of the Bittensor module. To run the subnet on the testnet, use the testnet/1.0.0 branch instead of main.