WER We Stand: Benchmarking Urdu ASR Models

TL;DR

This paper benchmarks Urdu ASR models using WER across read and conversational speech datasets, introduces a new conversational dataset, and highlights challenges in low-resource language ASR evaluation.

Contribution

It provides the first conversational speech dataset for Urdu ASR benchmarking and compares multiple models, revealing performance variations and evaluation challenges.

Findings

Seamless-large outperforms others on read speech

Whisper-large performs best on conversational speech

Evaluation of low-resource ASR models requires robust text normalization

Abstract

This paper presents a comprehensive evaluation of Urdu Automatic Speech Recognition (ASR) models. We analyze the performance of three ASR model families: Whisper, MMS, and Seamless-M4T using Word Error Rate (WER), along with a detailed examination of the most frequent wrong words and error types including insertions, deletions, and substitutions. Our analysis is conducted using two types of datasets, read speech and conversational speech. Notably, we present the first conversational speech dataset designed for benchmarking Urdu ASR models. We find that seamless-large outperforms other ASR models on the read speech dataset, while whisper-large performs best on the conversational speech dataset. Furthermore, this evaluation highlights the complexities of assessing ASR models for low-resource languages like Urdu using quantitative metrics alone and emphasizes the need for a robust Urdu text normalization system. Our findings contribute valuable insights for developing robust ASR systems for low-resource languages like Urdu.