AI Models: Accent Translation

Introduction

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

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

AT5.8

What changed?

• Upgraded to 16 kHz audio (up from 8 kHz), so voices sound crisper and closer to natural call quality.

• About 30% faster processing, so conversations feel more responsive in real time.

• A lighter model that uses about 45% less memory, which gives inbound deployments more headroom on lower-end agent machines.

• Automatic level control is now built into the voice model, which keeps output volume more consistent.

• More stable speaker identity, so the agent's voice stays consistent across a call.

• Clearer speech that's easier to understand, with better voice naturalness than the previous inbound model.

AT5.7

What changed?

• Upgraded to 16 kHz audio (up from 8 kHz) — voices come through louder, crisper, and closer to natural call quality.

• Reduced processing latency, so responses feel faster and more natural in real-time conversation.

• Automatic level control is now built directly into the voice model, keeping output volume more consistent.

• Reduced accent bleed for a cleaner, more consistent target-accent output.

• Significant improvement in voice naturalness and quality.

AT5

What changed?

• Source accents are no longer limited to specific countries and geographies. They are now universally available and applicable across nearly all regions worldwide.

• Real-time translation of incoming audio accents to improve comprehension. Inbound Accent Translation can be used independently or alongside Outbound Accent Translation.

• Headset-based Model Detection for Accent Translation — Sanas now automatically applies the optimal voice-isolation mode based on the connected headset:

  • Wired USB headsets → Near-field mode

  • Bluetooth or other devices → General mode

• For AT5.1, the following feature has been added:

  • Real-time translation of outgoing audio accents

AT4.8

What changed?

• Production-ready UK English model. Addresses the primary issue of flat, inexpressive synthesis in the previous British (UK-EN) model.

• Supports Source accent (India, Filipino, Latin American, African, and Middle Eastern).

• Improved intelligibility for Latin American and African/Middle Eastern Source accent via an updated Middleweight Translator (quantized for production efficiency).

• 8kHz output optimized for lightweight contact center deployments, paired with improved VAD for fewer hallucinations and voice breaks.

AT4.6

What changed?

• Significant intelligibility gains across all supported Source accent (India, Filipino, Latin American, African, and Middle Eastern), with the largest improvements for Middle Eastern and African English speakers.

• Optimized Voice Activity Detection reduces hallucinations and voice breaks during pauses.

• 8kHz output support for low-bandwidth telephony and contact center environments.

AT4.4

What changed?

• First UK English output model. Targets modern Received Pronunciation (RP) for broad intelligibility across the UK and European markets.

• Supports Indian, Filipino, Latin American, Middle Eastern, and African Source accent accents — validated through blind A/B evaluation with UK-based listeners who strongly preferred the RP output over the existing US model.

• Expanded African source accent coverage, trained on a proprietary dataset spanning 30+ subregional accents across North, East, West, and South Africa.

Info: Superseded by AT4.8. AT 4.8 addresses expressiveness and synthesis quality issues identified in this model. New deployments should use AT 4.8.

AT4.3

What changed?

• Significant intelligibility improvement for African and Middle Eastern Source accent over AT 4.1, reducing WER from 29.0% to 24.5% for this accent group.

• Finetuned on a Pan-African training dataset covering North, East, West, and South African regions, plus Caribbean accents.

• Uses an AT 4.0-style Middleweight Translator with 8kHz output and CPU utilization comparable to AT 2.8.

AT4.1

What changed?

• Major intelligibility improvement over AT 2.9 across Indian, Filipino, and Latin American English speakers, measurable in lower Word Error Rates (WER) for downstream ASR systems.

• Most natural-sounding AT model at time of release — improved intonation, pacing, and speaker similarity validated through large-scale blind A/B listening tests.

• New Lightweight mode runs with 35% fewer parameters, enabling deployment on thin clients and older hardware without sacrificing clarity.

Support