Indic DiarBench: A Multilingual Joint Diarization and ASR Benchmark for Indian Languages

Details

India's 22 scheduled languages represent one of the most linguistically diverse environments on earth, yet speech AI systems have been evaluated almost exclusively in English or other high-resource languages. Indic DiarBench addresses this with a benchmark spanning all 22 scheduled Indian languages. The dataset covers 108 hours of natural multi-speaker audio from near-field meetings, far-field recordings, and in-the-wild audio. All annotations are human-corrected with time-aligned, speaker-attributed transcriptions — capturing realistic conversational nuance including English code-mixing, dialectal variation, and frequent speaker overlap. These features are common in Indian speech but largely absent from existing benchmarks. The study evaluates leading systems including commercial speech APIs and multimodal large language models, establishing baselines for joint diarization and transcription. Indic DiarBench is released as an open-access resource to advance inclusive multilingual speech research. Accepted at Interspeech 2026, one of the top speech research venues, the benchmark provides a standardized and realistic evaluation environment for teams working on Indian-language ASR, speaker diarization, or code-switching.