Anuvani LID improves language identification for Indian speech
Anuvani-LID: Robust and Calibrated Spoken Language Identification for Indian Languages
Computation and Language
Summary
Language identification decides which language is spoken in audio before translating or processing it. Existing speech models struggle with many Indian languages, especially in short or noisy clips. The authors created Anuvani LID, a specialized model trained on a large collection of Indian speech, which identifies 22 Indian languages and Indian accented English more accurately and consistently than other models. Anuvani LID guesses the language better even on one-second clips and is carefully calibrated to measure its confidence well.
What this means in practice
- •For voice assistant developers: Integrate Anuvani LID to improve automatic detection of Indian languages in noisy or short speech segments for better user interaction.
- •For call center technology teams: Use Anuvani LID to accurately route calls to agents fluent in caller languages from 22 Indian languages and Indian accented English.$Commercial implications: Enables language-sensitive call routing products for Indian multilingual customer bases, improving service quality and operational efficiency.
Authors
Stephen Fernandes, Narayan Choudhary
Abstract
Every multilingual speech pipeline begins with a language decision, an error there propagates silently through recognition, translation and routing. In India it spans 22 scheduled languages across four language families, heard over short, noisy, narrowband and code-mixed audio. Foundational speech models increasingly present language identification (LID) as an inherent latent, generalized capability. Autoregressive decoders in speech models emit a special language ID token, and modern LLMs with audio capabilities can name the language when prompted. We test this assumption on identical clips and find that whisper and similar speech models has tokens for only 15 of the 23 languages, on one second inputs Gemini 3.5 Flash reaches 0.52 macro-F1 and GPT-audio 0.14. We present Anuvani LID, a dedicated 115M parameter LID conformer model trained on 54,643 hours of Indian speech spanning 22 Indian languages and Indian accented English. Along with designed and controlled studies on 102 condition benchmark with speaker and corpus held out tests. Across 37 conditions, it reaches 0.914 mean macro-F1, 5.2 points above Gemini 3.5 Flash and ahead of 14 other systems. It is 24 points ahead of Gemini at one second, scores 0.897 on all 23 languages where no other system exceeds 0.713, and is calibrated to an ECE of 0.008. Reliable Indian LID still requires a robust, dedicated and calibrated model.