Downloads · 30 days
0
mr00mr00/whisper-large-v3-turbo-fi-ggml
whisper-large-v3-turbo-fi-ggml is a automatic speech recognition model from mr00mr00. Use it when you need speech turned into text. It is set up for whisper.cpp. The card lists the license as apache-2.0.
GGML conversions of mozilla-ai/whisper-large-v3-turbo-fi, a Finnish fine-tune of OpenAI's whisper-large-v3-turbo, prepared for use with whisper.cpp and any tool that loads whisper.cpp GGML files (e.g. the Whispering d…
Downloads · 30 days
0
Access
Public
Updated May 18, 2026
Repo size
2.2 GB
Likes
0
Public
Click a slice to open those files.
.bin2.2 GB · 100%
From the Hugging Face model README
GGML conversions of mozilla-ai/whisper-large-v3-turbo-fi, a Finnish fine-tune of OpenAI's whisper-large-v3-turbo, prepared for use with whisper.cpp and any tool that loads whisper.cpp GGML files (e.g. the Whispering desktop transcription app, whisper-rs, etc.).
Published so others don't have to repeat the safetensors → GGML conversion.
| File | Format | Size | Notes |
|---|---|---|---|
ggml-large-v3-turbo-fi.bin | GGML f16 | 1.5 GB | Full-precision conversion of the source weights. Best quality, larger memory footprint. |
ggml-large-v3-turbo-fi-q5_0.bin | GGML q5_0 | 547 MB | Quantized version. Recommended for everyday use: ~3× smaller, faster load + inference, transcription quality essentially indistinguishable from the f16. |
Pick q5_0 first; switch to f16 only if you notice a meaningful quality regression on your audio.
./build/bin/whisper-cli -m /path/to/ggml-large-v3-turbo-fi-q5_0.bin -l fi -f audio.wav
use whisper_rs::{WhisperContext, WhisperContextParameters};
let ctx = WhisperContext::new_with_params(
"/path/to/ggml-large-v3-turbo-fi-q5_0.bin",
WhisperContextParameters::default(),
)?;
ggml-large-v3-turbo-fi-q5_0.bin to any folder, e.g.
~/Library/Application Support/com.bradenwong.whispering/models/whisper/ (macOS)
%APPDATA%/com.bradenwong.whispering/models/whisper/ (Windows)
~/.local/share/com.bradenwong.whispering/models/whisper/ (Linux)
.bin file.fi (or leave on auto).hf CLI downloadhf download mr00mr00/whisper-large-v3-turbo-fi-ggml \
ggml-large-v3-turbo-fi-q5_0.bin \
--local-dir .
Measured by the source repo on 1806 Finnish samples (Common Voice 22.0):
Baseline (whisper-large-v3-turbo) | Fine-tuned (this model) | |
|---|---|---|
| Word Error Rate (Normalized) | 16.576 | 11.972 |
| Word Error Rate (Orthographic) | 20.298 | 15.884 |
| Character Error Rate (Normalized) | 3.808 | 2.255 |
| Character Error Rate (Orthographic) | 4.498 | 3.119 |
GGML conversion is mathematically equivalent at f16. Quantization to q5_0 introduces a very small additional error that is empirically negligible on Whisper-class models; if you care about the last fraction of a percent of WER, use the f16 file.
Anyone can regenerate these files from scratch in about ten minutes plus model-download time. Recipe:
# Prereqs: git, cmake, uv (or any Python 3.10+ env manager).
# 1. Clone the conversion script's host repo and OpenAI's whisper repo
# (the script reads the mel filterbank assets from openai/whisper)
git clone https://github.com/ggerganov/whisper.cpp
git clone https://github.com/openai/whisper.git openai-whisper
# 2. Python env with conversion deps
cd whisper.cpp
uv venv --python 3.11
.venv/bin/uv pip install torch transformers numpy huggingface_hub
# 3. Download the source HuggingFace model (~3 GB)
.venv/bin/hf download mozilla-ai/whisper-large-v3-turbo-fi \
--local-dir ../whisper-large-v3-turbo-fi
# 4. Convert HF safetensors → GGML f16 (~30 s once the model is loaded)
mkdir -p converted
.venv/bin/python models/convert-h5-to-ggml.py \
../whisper-large-v3-turbo-fi \
../openai-whisper \
converted
# Output: converted/ggml-model.bin (1.5 GB, f16)
# 5. Build the quantize binary
cmake -B build -DBUILD_SHARED_LIBS=OFF
cmake --build build --target whisper-quantize -j
# 6. Quantize to q5_0 (~2 s on Apple Silicon)
./build/bin/whisper-quantize converted/ggml-model.bin converted/ggml-model-q5_0.bin q5_0
You can swap q5_0 for any other supported scheme listed by whisper-quantize --help (q4_0, q4_1, q5_1, q8_0, ...) to trade size vs. quality differently.
Licensed under the Apache License 2.0 (see LICENSE).
This is a derivative work; the upstream chain:
openai/whisper-large-v3-turbo MIT (base architecture & pretraining)
└── mozilla-ai/whisper-large-v3-turbo-fi Apache-2.0 (Finnish fine-tune)
└── this repo: GGML conversion + q5_0 Apache-2.0 (format conversion + quantization)
mozilla-ai/whisper-large-v3-turbo-fi)safetensors format to whisper.cpp's GGML binary format (f16).whisper-large-v3-turbo checkpoint.speech-to-text-finetune.If you use this model in a project, please cite the upstream fine-tune.