SoundsRight operates as a specialized subnet within the Bittensor decentralized ecosystem, focusing exclusively on speech enhancement technology. The subnet creates a competitive environment where participants (miners) develop and fine-tune speech enhancement models, which are then evaluated by validators to determine the best-performing solutions.

The core function of SoundsRight is to facilitate daily fine-tuning competitions for speech enhancement models. These competitions currently focus on two primary tasks:

• Denoising: Removing unwanted background noise from speech recordings while preserving the quality and intelligibility of the desired speech.
• Dereverberation: Reducing or eliminating the echo and reverberation effects that occur when audio is recorded in spaces with reflective surfaces.

Each competition follows a winner-takes-all format, which incentivizes miners to submit their absolute best models rather than multiple variations. This format, combined with the validation mechanism, deters miner factions by making model duplication unviable.

How SoundsRight Works

The SoundsRight subnet operates through a well-defined workflow involving miners, validators, HuggingFace (as a model repository), the Bittensor blockchain, and the subnet’s website. Here’s a detailed breakdown of how the system functions:

Miner-Validator Architecture

There are two main entities in the subnet:

1. Miners: These participants upload fine-tuned speech enhancement models to HuggingFace. Miners are responsible for developing and continuously improving speech enhancement models that can effectively perform denoising or dereverberation tasks.
2. Validators: These entities benchmark the models and determine which miners’ models perform best. Validators generate fresh datasets daily, download models from HuggingFace, verify model ownership, run benchmarks, and assign scores based on performance metrics.

Competition Workflow

The daily competition process follows these steps:

1. Miners fine-tune speech enhancement models and upload them to HuggingFace.
2. Validators generate new benchmarking datasets to ensure models are not susceptible to overfitting.
3. Validators send synapse requests for model information to the Bittensor chain.
4. The Bittensor chain returns synapse containing model information.
5. Validators reference model metadata and confirm model ownership.
6. Validators download models from HuggingFace.
7. Validators benchmark models on locally generated datasets.
8. Validators report benchmarking results to the subnet website.
9. The subnet website constructs competition leaderboards.
10. Validators set weights for miners based on performance.

This continuous cycle ensures that models are constantly being improved and evaluated on fresh data, driving innovation in speech enhancement technology.

Technical Architecture

The SoundsRight subnet is built on the Bittensor ecosystem, with a technical architecture designed to facilitate the competition and evaluation process efficiently.

Repository Structure

The codebase is organized into several key directories:

• soundsright/: Main code directory containing the core implementation
• base/: Contains base classes and utilities for the subnet
• core/: Core functionality of the subnet
• neurons/: Implementation of validator and miner neurons
• benchmarking/: Code for benchmarking models
• data/: Data handling utilities
• models/: Model definitions and implementations
• templates/: Template files
• utils/: Utility functions

Core Components

• BaseNeuron Class: Handles base operations for both miner and validator neurons.
• Validator Neurons: Responsible for benchmarking models and setting weights.
• Miner Neurons: Responsible for developing and uploading models.
• Configuration System: Manages paths, logging, and other parameters.

Technical Implementation

The implementation uses Python with dependencies including:

• argparse for command-line arguments
• bittensor for blockchain integration
• numpy for numerical operations
• Various file system and path handling utilities

Competition Metrics

The subnet currently hosts competitions at a 16 kHz sample rate, with plans to expand to 48 kHz competitions in upcoming updates. The benchmarking metrics used include:

• PESQ (Perceptual Evaluation of Speech Quality): 15% of total weights for denoising, 15% for dereverberation
• ESTOI (Extended Short-Time Objective Intelligibility): 12.5% of total weights for denoising, 12.5% for dereverberation
• SI-SDR (Scale-Invariant Signal-to-Distortion Ratio): 7.5% of total weights for denoising, 7.5% for dereverberation
• SI-SAR (Scale-Invariant Signal-to-Artifacts Ratio): 7.5% of total weights for denoising, 7.5% for dereverberation
• SI-SIR (Scale-Invariant Signal-to-Interference Ratio): 7.5% of total weights for denoising, 7.5% for dereverberation

These metrics ensure comprehensive evaluation of model performance across different aspects of speech enhancement quality.