Voice Isolation 2.5
AI model · Music & audio generation · Text → Audio / speech by Krisp
AI model that separates the primary speaker from background noise and secondary voices for Voice AI agents.
Voice Isolation 2.5 is a specialized AI model designed to separate a primary speaker's voice from background noise and secondary voices in real time. It is primarily intended for developers of Voice AI agents to improve the accuracy of downstream Speech-to-Text (STT) engines. The model operates with a low 15 ms latency and runs entirely on CPU, supporting high-concurrency server deployments. By removing cross-talk and background chatter, it reduces word error rates by an average of 46%. It is available in multiple flavors, including a "Lite" version that uses 70% less compute for edge and mobi
- Latency
- 15 ms
- Compute Requirement
- Runs on CPU
- Sampling Rate Support
- 16 kHz and 32 kHz
- WER Improvement
- 46% average reduction
- Optimization Support
- FP16 for Graviton3 and Sapphire Rapids
Overview
Voice Isolation 2.5 is a specialized AI model designed to separate a primary speaker's voice from background noise and secondary voices in real time. It is primarily intended for developers of Voice AI agents to improve the accuracy of downstream Speech-to-Text (STT) engines. The model operates with a low 15 ms latency and runs entirely on CPU, supporting high-concurrency server deployments. By removing cross-talk and background chatter, it reduces word error rates by an average of 46%. It is available in multiple flavors, including a "Lite" version that uses 70% less compute for edge and mobile applications.
Audio Enhancement
- 46% WER Reduction: Reduces word error rate (WER) by 46% on average by providing a clean signal to STT engines.
- 70% Background Speech Reduction: Achieves up to 70% reduction in word errors specifically in conditions with heavy background speech.
- Primary Speaker Separation: Separates the primary speaker from competing voices, background chatter, and room echo.
- 15 ms Latency: Operates with a 15 ms latency, making it suitable for real-time conversational AI.
Model Variants & Efficiency
- CPU Optimization: Supports 16-bit math (FP16) on AWS Graviton3 and Intel Sapphire Rapids to double live calls per server.
- HD Model Flavor: Designed for 32 kHz high-fidelity streams such as WebRTC or WebSocket.
- Default Model Flavor: Strongest model for extreme noise and badly degraded streams.
Pricing
- Voice Isolation API (Starter)
- $0.18 / hr
- Voice Isolation API (Growth)
- $0.12 / hr
Specifications
- Latency
- 15 ms
- Compute Requirement
- Runs on CPU
- Sampling Rate Support
- 16 kHz and 32 kHz
- WER Improvement
- 46% average reduction
- Optimization Support
- FP16 for Graviton3 and Sapphire Rapids
Added by AntSpace from public sources (krisp.ai).