語音 AI
AgenticASR Repeatedly Revises Transcripts With a Sliding Window, 1B Refiner Pushes Bilingual Benchmark Score to 79.95
AgenticASR separates speech recognition from text cleanup, allowing a small Refiner to rewrite a limited portion of existing output as new speech arrives, without repeatedly processing the full transcript. The code, checkpoints, and a 917-item Chinese-English benchmark are publicly available, but most of the audio is synthetically generated and scoring relies on an LLM judge.
Traditional streaming ASR aims for verbatim fidelity, typically retaining filler words, stutters, repetitions, and statements later retracted by the speaker. If each segment is then polished separately by a language model, the model may not see later cues that clarify or correct earlier speech. AgenticASR, released on July 30, uses a decoupled ASR–Refiner architecture: the front-end recognizer continuously produces text, while the Refiner merges the most recent source chunks, regenerates clean text, and directly replaces the corresponding section of the existing output.
By default, the system retains only three active chunks, so compute requirements depend on the sliding-window size rather than the total audio duration. The Refiner is initialized from MiniCPM-5-1B and fully fine-tuned on 100,000 synthetic spoken-text/written-text pairs. The researchers also pair it with Qwen3-ASR and three Whisper model sizes, showing that this layer does not need to be tied to a specific recognizer. The open-source streaming implementation integrates VAD, sherpa-onnx, stable chunking, and MLX-LM; its local backend currently focuses primarily on macOS.
The team also released AASR-Bench, which contains 917 Chinese and English samples and 6,637 atomic rules that separately evaluate content preservation, format normalization, filler filtering, and self-correction. Qwen3-ASR-1.7B paired with AgenticASR scored 79.95, outperforming post-processing with Qwen3.5-Flash at 69.93 and FormalASR, which directly generates formal text, at 52.50. Average end-to-end latency was 9.59 seconds, compared with roughly 60.89 seconds for the API post-processing baseline. The authors also report improvements across all five ASR front ends, although content preservation with weaker Whisper front ends still trailed the API baseline.
A key limitation is that most test audio was synthesized using TTS, with only 16.36% recorded by humans. The rules were generated by Qwen3.7-Plus and scored by Gemma-4-31B-IT, which may introduce model bias. Engineering teams should next validate the system on overlapping speech, accents, noise, and very long streams in real meetings, while also measuring product risks such as cursor jumps caused by transcript revisions or repeated triggering of downstream commands.