SmallALM

A lightweight 135M-parameter Audio Language Model for audio and speech understanding.

SmallALM is a lightweight Audio Language Model (ALM) with only 135M parameters, built around a simple and scalable design for audio and speech understanding.

The architecture combines four core components:

  • Audio Encoder — JASPER, a self-supervised encoder producing strong representations for both audio and speech signals.
  • Downsampler — average pooling with a factor of 8 to reduce the audio sequence length before it reaches the language model.
  • Projector — a two-layer MLP (768 → 576) with LayerNorm and GELU activation, bridging the audio encoder and the text decoder.
  • Text DecoderSmolLM2-135M, a small language model that generates text conditioned on the projected audio representations.

By pairing a frozen, self-supervised audio encoder with a compact SLM through a lightweight projector, SmallALM aims to show that competitive audio understanding doesn’t require billion-parameter models — making it practical to train and run on modest compute.

Code: github.com/hrishikeshhpillai/smallALM