Qwen3-ASR
3 storiesQwen3-ASR is a speech recognition model in the Qwen3 family. Recent coverage includes Nari Labs open-sourcing an inference engine for Qwen3-TTS and Qwen3-ASR, claiming sub-50 ms latency at 10 RPS; Hugging Face's AnySpeechRecognizer Swift package for Apple platforms, which lists Qwen3-ASR as a Core ML engine option; and a small study that used cocktail-party audio as an audio CAPTCHA, where Qwen3-ASR scored 0.55 word accuracy.
Related topics
Hugging Face released two open Swift packages for Apple platforms. AnySpeechRecognizer wraps Apple's iOS 26/macOS 26 SpeechAnalyzer API with selectable engines including Parakeet and Qwen3-ASR on Core ML, plus speaker diarization. AnyDecisionModel answers yes/no, multiple-choice and rating questions from probabilities rather than generated text. Both fetch models from the Hugging Face Hub and run on device.
A small study used 40 synthetic clips with background chatter to test whether speech models could serve as an audio CAPTCHA. Humans averaged 0.71 word accuracy and a 77% pass rate, versus 0.54 and 55% for instructed Gemini 3.8 Flash and 0.55 for Qwen3-ASR—too narrow a gap for a reliable test.
Nari Labs open-sourced an inference engine for Qwen3-TTS and Qwen3-ASR, claiming sub-50 ms latency at 10 RPS and top accuracy/cost performance on Coval benchmarks versus closed services; the team published the repository and reported outperforming some official endpoints.
That is everything
The A.I. Daily
Hugging Face launches AnySpeechRecognizer and AnyDecisionModel for Apple platforms
