AgenticASR: Refining Real-World Speech Recognition via Agentic Approaches and Benchmark
solyarisoftware · x · 2026-08-01
The official repository for AgenticASR has been released, aiming to optimize speech recognition (ASR) outputs in real-world scenarios using an agentic approach.
Key Features:
- Intent Preservation: Converts speech to clean written text by removing disfluencies and resolving self-corrections. It can even revise previously emitted text when later speech adds new evidence.
- Bilingual Support: Supports speech-to-clean-text refinement for both English and Chinese.
- ASR-Agnostic: The Refiner is decoupled from the recognizer and can be attached to any ASR frontend.
Released Assets: Includes the paper, source code, the AASR-Bench benchmark, and the Refiner checkpoint.
More from coding & agent
- Testing 105 Bugs: Luna Beats Fable in Cost-Efficiency, Expensive Models Avoid Coding — PawelHuryn · 2026-08-01
- Claude 3.5 Sonnet Generates 3D Game from Scratch Without Any Meshes or Assets — CtrlAltDwayne · 2026-08-01
- Acontext: Open-Source Skill Memory Layer for AI Agents — tom_doerr · 2026-08-01
- Stanford and GXL Host re:AGENT Hackathon for End-to-End Agentic Science — BrianHie · 2026-08-01
- Anthropic Engineers Shift from Prompting to Building Agentic Graphs — joemeno · 2026-08-01
- BaoCut Open-Sources Agent Skill: Let Claude Code Edit Videos via Natural Language — huangyun_122 · 2026-08-01