Clarity

Product Introduction

1. Definition: Clarity 1 is a real-time speech enhancement and target speaker extraction model, categorized as a streaming AI audio processing API. It is specifically engineered for voice agent and call center pipelines.
2. Core Value Proposition: It exists to solve the critical problem of ambient noise and overlapping speech in voice AI applications. Its primary function is to isolate the primary caller's voice by stripping out all background noise, chatter, and competing speakers in real time, ensuring voice bots and agents receive only clean, intelligible audio.

Main Features

1. Real-Time Noise Removal: Clarity 1 processes audio streams with ultra-low latency, removing non-speech background sounds like street noise, keyboard clicks, and café ambiance. It operates on 240 ms audio chunks, adding approximately 50 ms of processing latency, making it suitable for live, two-way conversation.
2. Primary Speaker Extraction (Target Speaker Extraction): This is the core differentiation. The model can identify and isolate a single target speaker from a mixture of voices. It uses a short reference audio sample of the target voice for precise extraction. Without a reference, it defaults to isolating the loudest speaker, effectively cutting out background conversations and cross-talk.
3. Streaming Architecture for Live Calls: Unlike post-processing tools, Clarity 1 is built for streaming. It processes audio continuously as it arrives, without waiting for a speaker to finish or for a full recording to complete. This is essential for real-time voice bot interactions and live call center audio feeds.

Problems Solved

1. Pain Point: Voice AI failure in noisy environments. Background noise and overlapping speech cause high word error rates (WER) in speech-to-text systems, false triggers in voice activity detection (VAD), and confused intent recognition in LLMs, leading to poor user experience and operational inefficiency.
2. Target Audience: Voice AI Developers and Engineers, Contact Center Technology Managers, Product Managers for conversational AI platforms, and System Integrators building voice-enabled solutions.
3. Use Cases: Essential for voice bots handling customer service calls from mobile phones in public spaces, call center agents working in open-plan offices, and any voice application where the caller's environment is unpredictable and noisy.

Unique Advantages

1. Differentiation: Clarity 1 combines both high-quality noise suppression and state-of-the-art target speaker extraction in a single, low-latency streaming model. Benchmarks show it outperforms or matches leading academic models (StarTSE) and production noise suppression tools (DeepFilterNet) in standard metrics like DNSMOS (SIG, BAK, OVRL) for both tasks.
2. Key Innovation: Its ability to perform streaming target speaker extraction with a 240 ms chunk size and ~50 ms latency is a significant technical advancement. Most research-level speaker extraction models are non-streaming, requiring full audio clips. Clarity 1 brings this capability to real-time production environments.

Frequently Asked Questions (FAQ)

1. How does real-time speech enhancement improve voice bot accuracy? By providing the speech-to-text (STT) engine and language model with a clean, single-speaker audio stream, Clarity 1 drastically reduces transcription errors caused by noise. This leads to more accurate intent understanding, fewer false turn-taking triggers, and more relevant bot responses.
2. What is the latency impact of adding Clarity to a voice pipeline? The total added latency is up to 290 ms. The model processes fixed 240 ms chunks. If your system feeds it smaller frames, the worst-case delay is 240 ms (waiting for the chunk to fill) + 50 ms (processing time). If your pipeline already uses 240 ms chunks, the added latency is only the 50 ms processing time.
3. Can Clarity 1 be used for post-processing recorded calls? While its architecture is optimized for streaming, the model can process pre-recorded audio files. However, its primary design and performance advantages are for real-time, low-latency applications like live voice AI and call center audio routing.
4. How do I provide a reference for the target speaker? For target speaker extraction, you provide a short, clean audio sample (a few seconds is sufficient) of the person you want the model to follow. This enrollment audio is used to create a voice profile. In scenarios like a call center, this could be the agent's voice at the start of a shift.
5. Is Clarity 1 an on-premise or cloud API solution? It is offered as a Cloud API, as indicated in its schema.org metadata. For large-scale or self-hosted deployments, the vendor (KugelAudio) invites direct contact, suggesting flexibility for enterprise needs.