November 5, 2026

Voice Arena introduces benchmark for speech recognition on phone calls

Voice Arena has launched a new benchmark for evaluating speech recognition, specifically focused on the complexities of two-person phone conversations, which include overlapping speech and interruptions. This initiative comes at a time when voice agents in support and sales face these real-world challenges daily, as they handle conversations filled with quick responses and audio compression inherent to phone lines. The first dataset, called the Monsoon Conversational ASR Test Set, is now available and features conversations in five languages, aiming to provide a more accurate assessment of speech-to-text models in authentic settings.

Voice Arena: Voice Arena is a platform dedicated to evaluating speech recognition models through specialized benchmarks. It recently launched a new benchmark specifically designed to test speech-to-text performance on authentic two-person phone conversations recorded in everyday environments with natural background sounds. The platform provides leaderboards and public test sets to support development across various languages and real-world use cases.
Shobhit Banga: Shobhit Banga is the individual who publicly introduced Voice Arena’s new speech recognition benchmark via announcement. His post highlighted the platform’s goal of creating one benchmark for every speech recognition use case across all languages, with an initial focus on real conversational datasets. He is directly associated with the launch of the Monsoon Conversational ASR Test Set.
Monsoon Conversational ASR Test Set: The Monsoon Conversational ASR Test Set is a publicly available dataset developed for testing automatic speech recognition systems. It features two-speaker conversations in five languages, captured on real phones in typical rooms with background noise and natural dialogue patterns including interruptions and short replies. This dataset forms the first part of Voice Arena’s broader benchmarking initiative for practical speech recognition scenarios.

Application Context: Voice agents in support and sales roles routinely encounter the exact conditions of natural, multi-turn phone dialogues that the new benchmark targets.
Evaluation Approach: Testing speech recognition on authentic everyday phone calls provides more relevant performance insights than traditional single-speaker or lab-based evaluations.
Real-world Challenges: Two-speaker phone conversations introduce significant difficulties for speech models through overlapping speech, interruptions, and audio compression from phone lines.

AI tools built by Emerald Force

Built and supported by Emerald Force.

You might also like