smol-audio
A tiernew this weekA free, open collection of Colab notebooks that makes fine-tuning Whisper, Parakeet, Voxtral, Granite Speech, and Audio Flamingo 3 actually approachable on commodity GPUs.
Kai's verdict
If you've ever rage-quit trying to fine-tune Whisper on a niche language or domain, smol-audio is the cookbook you wished existed — transparent, practical, and actually runs on free Colab. It's a practitioner's toolkit, not a product, but that's exactly what makes it useful. (Verdict pending Phi's full review.)
Strengths
- Covers five distinct state-of-the-art audio models in one repo — rare breadth for a single toolkit
- Designed to run on a standard 16 GB Colab T4 GPU, no local hardware needed
- Exposes full training loops and data pipelines transparently within the HuggingFace ecosystem (transformers, peft, accelerate, datasets)
- LoRA support baked in for memory-heavy models like Audio Flamingo 3 and Voxtral
- Apache 2.0 license — fully hackable and production-ready
Weaknesses
- No UI or web app — purely notebook-based, so non-developers need not apply
- Very new (released late April 2026), so community vetting, bug reports, and long-term maintenance are unproven
- Colab's free tier GPU availability is unreliable; longer fine-tuning runs may timeout or OOM without Colab Pro
Best for
ML engineers and audio researchers who want reproducible, low-friction recipes for fine-tuning open-source speech models on custom domains without standing up their own GPU infra.
Pricing
Free (open-source, Apache 2.0)
100% free; requires only a Google account for Colab. Colab Pro (~$10/mo) recommended for larger runs.