whisper-small-hu API & Inference Endpoint

Usage

python
from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="SubZtep/whisper-small-hu")
print(asr("audio.mp3", generate_kwargs={"language": "hungarian", "task": "transcribe"})["text"])

Training

Data: Common Voice 26.0 Hungarian train split (62,284 clips), silence-trimmed, 16 kHz
Schedule: 3 epochs (~11,700 steps), effective batch size 16 (8 × 2 gradient accumulation)
Optimizer: AdamW, lr 1e-5, linear decay with 500 warmup steps, fp16, gradient checkpointing
Hardware: single NVIDIA T4 (Google Colab)
Selection: best checkpoint by WER on a fixed 1,500-sample subset of the dev split

Training code: https://github.com/SubZtep/whisper-hu

Training progress

WER on the 1,500-sample dev subset during training:

Table with columns: Epoch, Step, Train Loss, Val Loss, WER
Epoch	Step	Train Loss	Val Loss	WER
0.26	1000	0.3302	0.3284	30.54
0.51	2000	0.2490	0.2674	25.36
0.77	3000	0.2130	0.2337	23.12
1.03

Framework versions

Transformers 4.44.2
PyTorch 2.11.0+cu128
Datasets 4.0.0
Tokenizers 0.19.1

Limitations

Trained on read speech (Common Voice sentences) — expect higher error rates on spontaneous conversation, noisy audio, or domain-specific vocabulary
WER is measured on lowercased text; the model itself outputs casing and punctuation, but their accuracy is not reflected in the score
Hungarian is agglutinative, so WER punishes single-suffix mistakes as whole-word errors — the CER of 3.33 is a better sense of how close transcripts are

Usage

python
from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="SubZtep/whisper-small-hu")
print(asr("audio.mp3", generate_kwargs={"language": "hungarian", "task": "transcribe"})["text"])

Training

Data: Common Voice 26.0 Hungarian train split (62,284 clips), silence-trimmed, 16 kHz
Schedule: 3 epochs (~11,700 steps), effective batch size 16 (8 × 2 gradient accumulation)
Optimizer: AdamW, lr 1e-5, linear decay with 500 warmup steps, fp16, gradient checkpointing
Hardware: single NVIDIA T4 (Google Colab)
Selection: best checkpoint by WER on a fixed 1,500-sample subset of the dev split

Training code: https://github.com/SubZtep/whisper-hu

Training progress

WER on the 1,500-sample dev subset during training:

Table with columns: Epoch, Step, Train Loss, Val Loss, WER
Epoch	Step	Train Loss	Val Loss	WER
0.26	1000	0.3302	0.3284	30.54
0.51	2000	0.2490	0.2674	25.36
0.77	3000	0.2130	0.2337	23.12
1.03

Framework versions

Transformers 4.44.2
PyTorch 2.11.0+cu128
Datasets 4.0.0
Tokenizers 0.19.1

Limitations

Trained on read speech (Common Voice sentences) — expect higher error rates on spontaneous conversation, noisy audio, or domain-specific vocabulary
WER is measured on lowercased text; the model itself outputs casing and punctuation, but their accuracy is not reflected in the score
Hungarian is agglutinative, so WER punishes single-suffix mistakes as whole-word errors — the CER of 3.33 is a better sense of how close transcripts are

whisper-small-hu

README

Usage

Training

Training progress

Framework versions

Limitations

Explore FriendliAI today

README

Usage

Training

Training progress

Framework versions

Limitations