Vividh-ASR: Malayalam Speech Recognition Test Set Stratified by Acoustic Complexity

Malayalam speech data from the Vividh-ASR benchmark, designed to evaluate automatic speech recognition models on real-world audio. The dataset is a test split organized by acoustic complexity to expose performance gaps in models trained predominantly on clean studio recordings. It was created by adalat-ai and last updated on Hugging Face in May 2026.

Use Cases

• Benchmarking ASR model robustness based on the described acoustic complexity stratification.
• Identifying model failure modes in real-world audio conditions as described in the benchmark's purpose.
• Evaluating the studio-bias of models fine-tuned on clean speech, a key concept mentioned in the description.

Strengths

• Designed specifically to evaluate ASR models on real-world, non-studio audio, addressing a stated gap in Indic ASR benchmarks.
• Organizes evaluation by acoustic complexity, a structured approach to exposing model weaknesses mentioned in the description.
• Last updated on 2026-05-20, suggesting recent maintenance.

Limitations

• Description metadata is limited; actual data quality, size, and file formats require manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and dataset scale are unknown, which may limit suitability assessment.

Provenance

Source

adalat-ai via Hugging Face.

Collection Method

Likely collected and curated for benchmarking purposes, but specific gathering method is not detailed.

Freshness

Last updated 2026-05-20 10:00:10.

Geography

Primarily Malayalam language, suggesting coverage of regions where Malayalam is spoken.