Dedicated model leads on Indian language speech tests

Language identification is often the first step in a speech system: it determines which recognizer or translation direction handles an utterance.

If general-purpose systems struggle with Indian languages, can a model built for them identify the language more reliably, including when the audio is short?

The authors introduce Anuvani-LID, a dedicated model that assigns one of 23 Indian-language labels to each speech utterance.

On identical clips in SUITE-42’s 37 core conditions, researchers averaged each system’s macro-F1 across conditions. Macro-F1 gives each language equal weight within a condition.

Anuvani-LID averaged 0.914 against Gemini 3.5 Flash’s 0.863, a 5.2-point difference. Anuvani-LID chose among 23 language labels, while Gemini received ten candidates.

This higher average supports the case for a dedicated model on these shared core tests. A separate five-condition panel tests all 23 languages.

Anuvani-LID, Voxlect Indic-Whisper, and Gemini were compared with all 23 labels available; their average macro-F1 scores were 0.897, 0.713, and 0.710, respectively. Some other systems cover fewer labels, and their scores count only languages they can emit.

The average advantage does not hold for every condition: Gemini was significantly ahead on six long, read-speech conditions. Anuvani-LID also assigns one label per utterance, covers only its 23 output classes, and was trained at full scale only once.