Hungarian Whisper small for FUTO Voice Input
I wanted good Hungarian dictation on my phone with FUTO Voice Input, which runs Whisper models fully offline. FUTO needs models in whisper.cpp's GGML format, so I took sarpba/whisper-hu-small-finetuned, a Hungarian fine-tune of OpenAI's Whisper small, and converted it.
That's all this is: the same model, repackaged. The hard work (the fine-tuning) is sarpba's, and the original model is OpenAI's. I didn't train anything.
Using it
On Android with FUTO Voice Input: download whisper-hu-small-finetuned_q8_0.bin, copy it to your phone, and
import it as a custom model in the app's settings.
With whisper.cpp:
whisper-cli -m whisper-hu-small-finetuned_q8_0.bin -f recording.wav -l hu
How well it works
I tested it on a Samsung Galaxy S23, and everyday Hungarian dictation worked flawlessly.
FUTO speeds things up by feeding Whisper only as much audio as you actually spoke instead of a full 30-second window. Some models start repeating themselves when you do that. This one didn't in my tests, and apart from one English word, it produced the same text both ways.
The weak spot is English words mixed into Hungarian. They tend to come out spelled the way they sound: when I said "coworking irodából", I got "Covercing irodából". If you dictate a lot of English terms, expect to fix those by hand.
For accuracy numbers, see the original model card. I haven't re-measured them, and the 8-bit quantization can shift results slightly.
How I made it
- Downloaded the original model from Hugging Face.
- Converted it to GGML with whisper.cpp's
convert-h5-to-ggml.py. - Quantized it to 8-bit (
q8_0) withwhisper-quantize, which shrinks it from ~970 MB to 252 MB.
To check your download:
whisper-hu-small-finetuned_q8_0.bin
size 264,464,624 bytes
sha256 b3bd92cb146181aef40dcdcdfa0af6fec181eacb499e22cfdefe20bdb1ec8bc2
Credits and license
- OpenAI Whisper: the original model (Apache-2.0)
- sarpba: the Hungarian fine-tune (Apache-2.0)
- whisper.cpp: conversion and quantization tools (MIT)
This conversion is released under Apache-2.0, like the models it comes from. It's a modified copy (converted and
quantized), as described in NOTICE. It isn't affiliated with or endorsed by FUTO, OpenAI or sarpba.