indian-languages

A team shares how they built a production voice-agent stack using self-hosted Pipecat, Mai Transcribe 2, a Sarvam conversational LLM, and a Malayalam fine-tuned VoxCPM2 TTS, achieving 20 concurrent low-latency voice streams on a single L4 GPU with an automated WhatsApp post-call handoff.

This paper presents COILD, an Indic-centric parallel corpus with over 1.16 million sentence pairs for 20 Indian language pairs and a domain-centric benchmark, demonstrating improvements in machine translation when fine-tuning multilingual models.

This paper investigates the structural sensitivity of multilingual large language models to semantics-preserving perturbations in Hindi and Malayalam, showing significant degradation in mathematical reasoning performance and introducing the IndicReStruct benchmark for evaluation.

The paper introduces INCLUDE, a multilingual evaluation benchmark to quantify Indian-centric socio-cultural biases in LLMs, revealing that non-English Indian languages exhibit higher bias than English, indicating cross-lingual safety alignment gaps.

Cartesia updates its voice generation model Sonic 3.6 with improved naturalness and accuracy for Indian languages, including Hindi and Hinglish, and adds support for Urdu and Odia.

Indic DiarBench is a multilingual joint diarization and ASR benchmark covering all 22 scheduled languages of India with 108 hours of human-corrected multi-speaker audio, capturing conversational nuances like code-mixing and speaker overlap.

This paper introduces BHARATI, a set of SentencePiece BPE tokenizers trained on a balanced multilingual corpus for classical Indian languages. Subword fertility analysis shows significant improvements in tokenization efficiency, reducing sequence length by up to 90% compared to baseline tokenizers, thereby enhancing effective context length for downstream language models.

This paper proposes a multimodal framework that jointly improves Automatic Speech Recognition (ASR) and Dialect Identification (DID) for Indian languages, using a Bottleneck Encoder and RoBERTa with a gating mechanism. Evaluated on eight languages with 33 dialects, it achieves 81.63% DID accuracy and reduces CER/WER to 4.65%/17.73%.

This paper audits multilingual clinical ASR systems on psychiatric interviews in Indian languages and proposes SamaVaani, a unified debiasing technique to improve performance and fairness across demographic groups.

OpenAI introduced IndQA, a new benchmark with 2,278 questions across 12 Indian languages and 10 cultural domains, designed to evaluate AI models' understanding of culturally nuanced and reasoning-heavy tasks that existing benchmarks fail to capture. Created with 261 domain experts, IndQA addresses the saturation of existing multilingual benchmarks like MMMLU and focuses on real-world cultural comprehension rather than translation or multiple-choice tasks.