General-purpose large language models (LLMs) and foundational speech models struggle with basic language identification for Indian languages, according to new research from Stephen Fernandes and Narayan Choudhary. Their paper, submitted on October 14, 2026, found that widely used models either lack support for many Indian languages or perform poorly on short, noisy audio clips. The paper shows a persistent need for specialized resources given India’s 22 scheduled languages.

Every multilingual speech processing pipeline relies on an accurate initial language decision. Errors made at this stage can propagate, affecting subsequent speech recognition, translation, and content routing. While modern speech models often present language identification (LID) as an inherent capability, Fernandes and Choudhary tested this assumption directly. They found that popular models such as Whisper only include specific language ID tokens for 15 out of 23 languages. When tested on one-second audio inputs, Gemini 3.5 Flash achieved a macro-F1 score of 0.52, while GPT-audio recorded 0.14. These figures indicate significant limitations for practical applications involving diverse Indian speech.

To address this, the authors present Anuvani-LID, a dedicated 115 million parameter Conformer model. This model was trained on 54,643 hours of Indian speech data, encompassing 22 Indian languages and various Indian-accented English speech. The dataset includes audio that is short, noisy, narrowband, and code-mixed, reflecting real-world conditions. Anuvani-LID aims to provide the accurate and calibrated language identification needed for pipelines handling India’s linguistic diversity across four language families.

Carefully constructed datasets and evaluations also matter for text-based applications. Gaurab Baral, Mimansak Nepal, and Grishma Basnet, in their paper submitted on October 13, 2026, introduced NEPSEE, a new benchmark for Nepali sentiment analysis. Their work emphasizes the common issue of single-annotator bias in sentiment datasets, where test labels often reflect the conventions of one person.

NEPSEE tackles this by employing a double-annotation process. Two annotators independently labeled 5,614 sentences from tweets by Nepali public figures on a five-point sentiment scale. For the three-class benchmark (positive, negative, neutral), only the 3,189 sentences where both annotators agreed on the polarity were kept. On this three-class task, NepBERTa with a fixed tokenizer achieved a macro-F1 score of 0.877, and IndicBERTv2 reached 0.867. Three commercial large language models performed slightly better, ranging from 0.882 to 0.896 in zero-shot settings, though the authors note these differences were not statistically significant.

For a more granular five-point sentiment scale, the researchers identified 766 sentences where annotators provided adjacent labels (for example, “Positive” and “Slightly Positive”). Both labels were accepted for training on this task. Here, Claude Opus 5.5 outperformed the best encoder models by 0.10 to 0.12 macro-F1, indicating that larger, more generalized models can show an advantage on finer-grained sentiment tasks when the data is carefully prepared.

A secondary finding from the NEPSEE paper shows a problem in fundamental text processing. A tokenizer audit revealed that the default tokenizers of three Nepali BERT models delete non-spacing Devanagari characters. This finding echoes previous concerns about fundamental components for Indic languages. This publication reported on October 14, 2026, that research into Tamil-first models also found tokenizer efficiency to be a key area for improvement. The recurring problem of tokenizers mishandling script-specific nuances suggests that foundational text processing layers require dedicated attention for Indian and related languages.

These two papers, one on speech identification and the other on text sentiment and tokenization for Nepali, show that general-purpose AI models often reveal significant gaps when applied to the specific challenges of Indian languages and scripts. Specialized models like Anuvani-LID are essential for reliable AI systems in this domain. So are high-quality, double-annotated benchmarks like NEPSEE, and the identification of core issues like tokenizer failures. Such efforts align with a broader pattern of Indian AI research developing new benchmarks for specialized tasks, as this outlet covered on September 29, 2026, when reporting on benchmarks for bias, conversational AI, and word similarity.

Sources

Compiled by Launch91 Desk from the sources linked above. More about Launch91.