Anuvani-LID: Robust and Calibrated Spoken Language Identification for Indian Languages
Not on the map. Why?
Abstract (arXiv)
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.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 49 of its 126 references that are in the map. First posted on arXiv 14 Oct 2026.
Builds on
49 of 126 references.
References · 49
- 2024The Llama 3 Herd of Modelscited by 14,715
- 2015Distilling the Knowledge in a Neural Networkcited by 10,324
- 2015Rethinking the Inception Architecture for Computer Visioncited by 5,475
- 2017On Calibration of Modern Neural Networkscited by 4,960
- 2019Parameter-Efficient Transfer Learning for NLPcited by 4,236
- 2022Robust Speech Recognition via Large-Scale Weak Supervisioncited by 4,057
- 2019Unsupervised Cross-lingual Representation Learning at Scalecited by 3,935
- 2020wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representationscited by 3,905
- 2016A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networkscited by 2,385
- 2020Shortcut Learning in Deep Neural Networkscited by 1,969
- 2020Conformer: Convolution-augmented Transformer for Speech Recognitioncited by 1,863
- 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,396
37 more
- 2019Common Voice: A Massively-Multilingual Speech Corpuscited by 1,384
- 2019SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognitioncited by 1,380
- 2018Averaging Weights Leads to Wider Optima and Better Generalizationcited by 1,209
- 2015NetVLAD: CNN architecture for weakly supervised place recognitioncited by 941
- 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 876
- 2023Large Language Modelscited by 780
- 2021SUPERB: Speech processing Universal PERformance Benchmarkcited by 666
- 2020ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verificationcited by 628
- 2020Long-tail learning via logit adjustmentcited by 618
- 2021XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scalecited by 601
- 2022FLEURS: Few-shot Learning Evaluation of Universal Representations of Speechcited by 469
- 2021Dynabench: Rethinking Benchmarking in NLPcited by 459
- 2015Deep Speech 2: End-to-End Speech Recognition in English and Mandarincited by 451
- 2015MUSAN: A Music, Speech, and Noise Corpuscited by 425
- 2023Scaling Speech Technology to 1,000+ Languagescited by 411
- 2020Calibrating Deep Neural Networks using Focal Losscited by 357
- 2021W2v-BERT: Combining Contrastive Learning and Masked Language Modeling for Self-Supervised Speech Pre-Trainingcited by 347
- 2021Layer-wise Analysis of a Self-supervised Speech Representation Modelcited by 298
- 2018Additive Margin Softmax for Face Verificationcited by 278
- 2023LEACE: Perfect linear concept erasure in closed formcited by 267
- 2019Massively Multilingual Neural Machine Translation in the Wild: Findings and Challengescited by 257
- 2019Competence-based Curriculum Learning for Neural Machine Translationcited by 202
- 2020VoxLingua107: a Dataset for Spoken Language Recognitioncited by 122
- 2018Attentive Statistics Pooling for Deep Speaker Embeddingcited by 114
- 2021Multilingual and code-switching ASR challenges for low resource Indian languagescited by 53
- 2025mmBERT: A Modern Multilingual Encoder with Annealed Language Learningcited by 45
- 2024IndicVoices: Towards building an Inclusive Multilingual Speech Dataset for Indian Languagescited by 41
- 2022IndicSUPERB: A Speech Processing Universal Performance Benchmark for Indian languagescited by 40
- 2023Svarah: Evaluating English ASR Systems on Indian Accentscited by 16
- 2019Self Multi-Head Attention for Speaker Recognitioncited by 13
- 2023SPRING-INX: A Multilingual Indian Language Speech Corpus by SPRING Lab, IIT Madrascited by 11
- 2024LAHAJA: A Robust Multi-accent Benchmark for Evaluating Hindi ASR Systemscited by 10
- 2022Accidental Learners: Spoken Language Identification in Multilingual Self-Supervised Modelscited by 6
- 2022A Compact End-to-End Model with Local and Global Context for Spoken Language Identificationcited by 5
- 2021Cross-Corpora Language Recognition: A Preliminary Investigation with Indian Languagescited by 4
- 2023Cross-Corpora Spoken Language Identification with Domain Diversification and Generalizationcited by 4
- 2026A Comparative Study of Pre-trained Speech Encoders and Training Objectives for Large-Scale Indic Spoken Language Identificationcited by 1
Built on by
Not cited yet.