[KongchangAI]
ProductSada-ASR

Sada-ASR

A production-grade speech recognition system for Arabic dialects, developed by EchoNova, intended to address the performance degradation of Modern Standard Arabic recognizers in dialect scenarios.

Timeline (last 90 days)

Nov 17

Sada-ASR is a production-grade Arabic voice assistant/speech recognition system developed by EchoNova.

Unverified50%
Nov 17

The Sada-ASR team built a 120-hour test set covering MSA and eight Arabic dialects.

Unverified50%
Nov 17

On the 120-hour test set, Sada-ASR achieved an overall word error rate (WER) of 11.69%, while the strongest comparison system, Microsoft MAI-Transcribe-2, achieved 13.51%.

Unverified50%
Nov 17

On segments unaffected by digits and writing conventions, Sada-ASR's WER lead over Microsoft MAI-Transcribe-2 was 0.74 percentage points.

Unverified50%
Nov 17

Sada-ASR's average word error rate across dialects was 15.52%, compared with an average of 16.96% for the other comparison systems.

Unverified50%
Nov 17

Sada-ASR leads competitors by 4.30 percentage points on Omani dialect, 3.78 percentage points on Egyptian dialect, and 2.19 percentage points on Saudi dialect.

Unverified50%
Nov 17

On Lebanese dialect, Sada-ASR's word error rate was 24.96%, trailing the competitor's 23.05%.

Unverified50%
Nov 17

The Sada-ASR team explains that the main reason for the weaker performance on Lebanese dialect is that it transcribes Lebanese vocabulary into MSA forms, resulting in lower verbatim transcription accuracy.

Unverified50%
Nov 17

The Sada-ASR team open-sourced a 20-hour subset of the test set along with the scoring code.

Unverified50%

All Facts (9)

Source Articles