GRPO treats ASR pseudo-labels as rewards rather than targets
In banking audio, GRPO cut WER 29.3% to 56.2% relative to SFT and copied 27% of teacher errors.
Stories per day · 30 days
3 RSS
Relatedspeech-to-text2ai-agents1benchmark1cactus-whistle1cpu1grpo1model-backdoor1pseudo-labeling1reinforcement-learning1speech-recognition1voice-trigger1
9 stories, newest first · one card per story
Stories tagged #whisper
Speech-model backdoor fires on one speaker’s voice
In sandbox tests, a poisoned transcript asking for $5,000 led some GPT-4o and Claude agents to pay.
Cactus Whistle, a 16.9 MB speech model, tested against Whisper Base
The guide is framed around word error rate, speed, streaming, timestamps, keywords and speech-to-tool calls on CPU.
3
OpenPA: Forced-alignment-free Multi-granular Pronunciation Assessment with Mixed-transcript Adaptation
OpenPA uses Whisper cross-attention to score pronunciation at word and utterance level without forced alignment.
Does Financial Speech Need Continual Learning Regularization? A Study of EWC and mini-KFAC on Earnings Call ASR
Adapting Whisper-small to Earnings22 improves financial ASR, so EWC regularization is often counterproductive.
Speech event detector: Timestamped detection of non-verbal vocalizations and background sounds for rich speech transcription
A 5.8M model timestamps 54 non-speech sounds and merges them into Whisper transcripts.
Itgan at NADI 2026 shared task: Parameter-Efficient Whisper Adaptation for Robust, Mixed-Dialect and Code-Switched Arabic ASR
LoRA-adapted Whisper systems for NADI 2026 Arabic ASR reach 57.1%, 46.7%, and 14.49% WER on three subtasks.
Show HN: Parrot – Open-Source Smart Meeting Recorder with Co-Pilot on Mac
Parrot 0.24.2 is an open-source Mac meeting recorder that suggests answers from local documents during calls.
Six Layers Less: Encoder Pruning for Whisper with Label-Free Recovery
Six Whisper encoder layers are pruned and partly recovered with label-free distillation.