mahwizzzz/moonshine-tiny-ur

Moonshine Tiny Ur

Urdu speech recognition at Moonshine Tiny size. The weights are a fine-tune of UsefulSensors/moonshine-tiny (27M parameters, MIT). The architecture is Moonshine (Jeffries et al., 2024, arXiv:2410.15608). The training recipe follows Flavors of Moonshine (King et al., 2025, arXiv:2509.02523): schedule-free AdamW, learning rate 2e-5, and an effective batch of 32.

Data

Source	Role	License	Clips	Hours
Audiobook speech, same clip count as mahwizzzz/1DamnAudio	real	cc-by-4.0	220100	189.98
mahwizzzz/sig, female and male	synthetic	mit	35464	61.64
humairawan/UrduSpeech, Urdu train	real	cc-by-4.0	23864	42.12
Common Voice Corpus 22.0 Urdu train	real	cc0-1.0	7188	8.29
google/fleurs, ur_pk train	real	cc-by-4.0	1962	6.18

Evaluation

Limitations

There is no pseudo-labeled radio corpus. Synthetic speech is capped, and the real pool is mostly read, studio, and audiobook speech.

License

CC-BY-4.0, because FLEURS, UrduSpeech, and the audiobook source are CC-BY-4.0. The upstream Moonshine Tiny checkpoint is MIT. Common Voice is CC0.

Usage

from transformers import AutoProcessor, MoonshineForConditionalGeneration

repo = "mahwizzzz/moonshine-tiny-ur"
processor = AutoProcessor.from_pretrained(repo)
model = MoonshineForConditionalGeneration.from_pretrained(repo)
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
ids = model.generate(**inputs, num_beams=1)
text = processor.batch_decode(ids, skip_special_tokens=True)[0]