[✨ FEATURE] New ASR model

[dxcore35]

https://www.youtube.com/watch?v=N9Z6-EvSYF8

https://github.com/Blaizzy/mlx-audio - native support for MLX
https://huggingface.co/models?search=Qwen-3-ASR

Some discusion with comparision with NVIDIA model:
https://x.com/mpuig/status/2017681772872396876?s=20

[altic-dev]

Interesting. I thought qwen wasnt the best yet. I will add it in after a quick testing :))

[dxcore35]

Category	Parameter / Feature	Description
Primary Output	Multilingual Text	Transcribes speech in 30 major languages and 22 Chinese dialects.
Time & Sync	Word-Level Timestamps	Extracts precise start/end times for every word or Chinese character.
Language	Language ID (LID)	Automatically identifies which of the 52 supported languages/dialects is being spoken.
Language	Accent Recognition	Recognizes English accents from 16 different global regions.
Emotion	Emotional State	Can identify and describe the speaker's mood (Happy, Sad, Angry, etc.) via prompting.
Tone	Prosody & Style	Extracts vocal characteristics like sarcasm, urgency, hesitation, or whispering.
Audio Scene	Audio Type Classification	Differentiates between normal speech, Singing, and Songs with BGM (Background Music).
Robustness	Noise Filtering	Extracts clean text even in "low SNR" (High Noise) or complex acoustic environments.
Diarization	Speaker Attribution	Identifies transitions between different speakers in a conversation.
Reasoning	"Thinking" Logic	In "Thinking Mode," the model can reason about slurred or ambiguous audio before transcribing.

Emotion can be interesting for your later AI prompt improvement to adjust prompt with that :)