Skip to content

Model Downloader

The Model Downloader fetches Whisper GGUF models directly from HuggingFace into your project's Plugins/AbolethSTT/Models/ directory.

Access via:

Project Settings > Plugins > Aboleth Speech-to-Text > Model Management


Model Size Speed Accuracy Notes
Tiny English ~75 MB Fastest Low Fastest model, lowest accuracy. Good for prototyping or low-spec hardware
Large V3 Turbo Q5 ~1.1 GB Fast Excellent Recommended -- best balance of size, speed, and accuracy
Large V3 Turbo Q8 ~1.6 GB Fast Excellent Near-lossless quantization, slightly larger than Q5
Large V3 Turbo ~3.1 GB Fast Excellent Full-precision Turbo, multilingual

Recommended Model

ggml-large-v3-turbo-q5_0.bin is the best general-purpose choice. It delivers excellent accuracy at one-third the file size of the full-precision Turbo model, with minimal latency overhead from quantization.


Turbo Architecture

The Large V3 Turbo models use a distilled architecture that is significantly faster than the older Small and Medium models despite being larger in file size. This is because the Turbo encoder is shallower and more efficient -- it was distilled from the full Large V3 model to retain accuracy while reducing compute.

Quantized variants (Q5, Q8) add negligible latency compared to full precision while reducing file size by up to 3x.


VRAM Usage

VRAM consumption during inference approximately equals the model's file size on disk.

Model Approximate VRAM
Tiny English ~75 MB
Large V3 Turbo Q5 ~1.1 GB
Large V3 Turbo Q8 ~1.6 GB
Large V3 Turbo ~3.1 GB

Warning

Choose a model that fits within your target hardware's available VRAM. If running alongside other GPU-intensive systems (rendering, other ML models), account for their VRAM usage as well.


Download & Storage

Models are downloaded from HuggingFace and stored in:

Plugins/AbolethSTT/Models/

The Silero VAD model (ggml-silero-v6.2.0.bin) is also available through the downloader, though it ships bundled with the plugin by default.

Note

The Model Downloader requires an internet connection. For air-gapped environments, download models manually from HuggingFace and place them in the Models/ directory.