Tiny Audio

Speech recognition combining Whisper encoder with SmolLM3 decoder.

Usage

from transformers import AutoModel

model = AutoModel.from_pretrained("mazesmazes/tiny-audio", trust_remote_code=True)
transcription = model.transcribe("audio.wav")

Architecture

  • Encoder: Whisper-small (frozen)
  • Projector: RMSNorm → Linear projection → AvgPool (2x downsampling) → RMSNorm
  • Decoder: SmolLM3 with LoRA

Training

Datasets: LibriSpeech, GigaSpeech, Common Voice, LoquaciousSet

  • BF16 mixed precision
  • Streaming datasets
  • Frozen encoder, LoRA fine-tuning on decoder

Links

Downloads last month
968
Safetensors
Model size
3.72B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for mazesmazes/tiny-audio

Adapter
(3)
this model

Datasets used to train mazesmazes/tiny-audio

Space using mazesmazes/tiny-audio 1