Research map / Large language models / Language, safety and interpretability
Towards Building ASR Systems for the Next Billion Users
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pretrain rather than train the acoustic model from scratch (ablation) | IndicWav2Vec BASE · MSR Gujarati | WER: 46.0 -> 22.8 | Helped |
| use the larger, more diverse pretraining corpus rather than EkStep pretraining data (ablation) | IndicWav2Vec BASE · MUCS Gujarati | WER: 30.3 -> 29.4 | Helped |
| use LARGE rather than BASE (ablation) | IndicWav2Vec · MUCS Gujarati | WER: 29.4 -> 26.2 | Helped |
| add a transcript-trained language model (ablation) | IndicWav2Vec M3 · MUCS Gujarati | WER: 26.2 -> 18.0 | Helped |
| train the language model on IndicCorp as well as transcripts (ablation) | IndicWav2Vec M4 · MUCS Gujarati | WER: 18.0 -> 17.2 | Helped |
| augment the task lexicon with frequent IndicCorp words (ablation) | IndicWav2Vec M5 · MUCS Gujarati | WER: 17.2 -> 14.8 | Helped |
| augment the task lexicon with frequent IndicCorp words (ablation) | IndicWav2Vec M5 · MSR Gujarati | WER: 11.7 -> 12.3 | Hurt |
| rescore hypotheses with an external Transformer language model (ablation) | IndicWav2Vec M6 · MUCS Gujarati | WER: 14.8 -> 14.3 | Helped |
| use multilingual fine-tuning with one shared softmax (ablation) | IndicWav2Vec M6 · MUCS Marathi | WER: 12.2 -> 30.3 | Hurt |
| use multilingual fine-tuning with separate per-language softmax layers instead of one shared softmax (ablation) | IndicWav2Vec · MUCS Marathi | WER: 30.3 -> 11.7 | Helped |
| add iterative pseudo-labeling (ablation) | IndicWav2Vec M6 · MUCS Gujarati | WER: 14.8 -> 15.1 | Hurt |
| increase fine-tuning data from 1 hour to 40 hours (ablation) | IndicWav2Vec · MSR Gujarati | WER: 35.9 -> 20.9 | Helped |
Closest papers by meaning
- IndicVoices-R: Unlocking a Massive Multilingual Multi-speaker Speech Corpus for Scaling Indian TTS (NeurIPS 2024 · similarity 0.90)
- LAMA-UT: Language Agnostic Multilingual ASR Through Orthography Unification and Language-Specific Transliteration (AAAI 2025 · similarity 0.89)
- WorldSpeech: A Multilingual Speech Corpus from Around the World (NeurIPS 2026 · similarity 0.89)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.88)
- VAANI: Capturing the language landscape for an inclusive digital India (NeurIPS 2026 · similarity 0.88)
- SwitchLingua: The First Large-Scale Multilingual and Multi-Ethnic Code-Switching Dataset (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.