Research map / Large language models / Language, safety and interpretability
Speech-to-LaTeX: New Models and Datasets for Converting Spoken Equations and Sentences
Dmitrii Korzh, Dmitrii Tarasov, Artyom Iudin, Elvir Karimov, Matvey Skripkin, Nikita Kuzmin, Andrey Kuznetsov, Oleg Rogov, Ivan Oseledets
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| train on human audio instead of TTS-only audio (ablation) | Qwen2.5 post-correction · S2L-sentences (human test, sentence CER) | CER: 33.74 -> 29.18 (Q-0.5B) | Improved |
| add 400k MathBridge TTS samples to training (Mix-full) (ablation) | Qwen2.5-0.5B post-correction · S2L-equations (English test, mix) | CER: 32.33 -> 27.21 | Improved |
| add 400k English TTS samples (ablation) | post-correction models · S2L-equations (Russian test) | CER | Hurt |
| add Russian data to training (ablation) | Qwen2.5-0.5B (human train) · S2L-equations (English human test) | CER: 35.01 -> 33.94 (0.5B); 27.57 -> 28.85 (1.5B) | Mixed |
| use math-specialized Qwen2.5-Math-1.5B instead of Qwen2.5-1.5B (ablation) | post-correction · S2L-equations (English) | CER | No clear effect |
| add LaTeX symbols as extra tokens to tokenizer (ablation) | post-correction LLMs · S2L-equations | CER | No clear effect |
| normalize LaTeX with AST-based KaTeX fork | evaluation of S2L-equations · S2L-equations | CER: reduced CER by ~1% | Improved |
| use end-to-end SALMONN audio-LLM instead of ASR post-correction | S2L-equations · S2L-equations (English test, mix) | CER: 25.69 -> 17.50 | Improved |
| use few-shot prompting instead of fine-tuning (ablation) | Qwen2.5 models · S2L-sentences | Equation CER: 43.75 -> 57.91 (Q-7B, 5-shot) | Hurt |
| scale post-correction LLM to 7B with LoRA (ablation) | Qwen2.5 post-correction · S2L-equations (English) | CER | No clear effect |
| scale Qwen2.5 to 7B with LoRA (ablation) | Qwen2.5 post-correction · S2L-sentences | Equation CER: 54.07 -> 43.75 (0.5B Mix vs 7B Mix) | Improved |
| add TTS audio to human audio for training (ablation) | SALMONN-13B · S2L-sentences (human test) | Equation CER: 41.53 -> 39.68 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MathSpeech — MathSpeech: Leveraging Small LMs for Accurate Conversion in Mathematical Speech-to-Formula (AAAI 2025) · better on 2 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MathSpeech: Leveraging Small LMs for Accurate Conversion in Mathematical Speech-to-Formula (AAAI 2025 · similarity 0.93)
- Exploring Equation as a Better Intermediate Meaning Representation for Numerical Reasoning of Large Language Models (AAAI 2024 · similarity 0.84)
- A Compact Model for Mathematics Problem Representations Distilled from BERT (AAAI 2025 · similarity 0.84)
- Recording for Eyes, Not Echoing to Ears: Contextualized Spoken-to-Written Conversion of ASR Transcripts (AAAI 2025 · similarity 0.84)
- MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark (ICLR 2026 · similarity 0.83)
- ATLAS: Autoformalizing Theorems through Lifting, Augmentation, and Synthesis of Data (NeurIPS 2025 · similarity 0.83)
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.
