AI models: Speech Enhancement

Introduction

The following document lists Sanas' Speech Enhancement (SE) model releases. Versions are listed newest-first. Each entry summarizes what changed and who it affects.

Speech Enhancement builds on all Noise Cancellation capabilities and adds high-fidelity audio processing at a higher sampling rate, producing richer, more natural-sounding output. Like NC, SE uses independently trained AI models for each audio direction— outbound (agent-side) and inbound (customer-side).

For desktop application releases that bundle these models, see Product updates. For detailed benchmarking methodology and performance data, see Sanas science blogs.

SE2.1

What changed

• Eliminates reverbs and room acoustics from speech

• Restores audio packet loss due to a poor network connection

• Corrects audio degradation via the codec compression

• Improved audio fidelity on the old telephony systems (8kHz)

• Improvements in the background environmental noises and distant speech removal

• Overall enhancements in speech clarity, articulation, energy, and vocal presence.

SE1.0.1

What changed

• Includes all outbound Noise Cancellation capabilities, background noise suppression up to 80–90 dBA, background speech cancellation, and support for 8,000+ noise profiles.

• High-fidelity 24kHz audio output delivers richer, more natural-sounding speech compared to the 8kHz telephony-grade output of NC models.

SE1.0.1

What changed

• Includes all inbound Noise Cancellation capabilities — far-field speech preservation, handling of unpredictable ambient environments, and support for variable-quality audio sources.

• High-fidelity 24kHz audio output enhances the clarity and naturalness of the customer's voice beyond what NC inbound models provide.

Support