language:
- mn
license: apache-2.0
library_name: transformers
pipeline_tag: automatic-speech-recognition
tags:
- whisper
- automatic-speech-recognition
- mongolian
- speech
- audio
base_model: openai/whisper-small
datasets:
- Blgn94/mongolian-stt-dataset-v24
metrics:
- wer
- cer
model-index:
- name: whisper-small-mn-v24
results:
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
name: Blgn94/mongolian-stt-dataset-v24 (validation)
type: Blgn94/mongolian-stt-dataset-v24
split: validation
metrics:
- type: wer
value: 15.16
name: WER
- type: cer
value: 6.6
name: CER
whisper-small-mn-v24
Fine-tune of openai/whisper-small
for Mongolian (mn) automatic speech recognition, trained on
Blgn94/mongolian-stt-dataset-v24.
This is the v24 iteration and improves substantially over the previous release
Blgn94/whisper-small-mn.
Results
| Metric | v24 (this model) | Previous whisper-small-mn |
|---|---|---|
| WER | 15.16% | 22.1% |
| CER | 6.60% | 10.3% |
That is a ~31% relative WER reduction and ~36% relative CER reduction over the previous model.
Note on measurement: these numbers are from a fixed 200-sample subset of the validation split, evaluated with
predict_with_generateduring training (the same protocol used for the previous model, so the comparison is apples-to-apples). A full 2,297-row validation evaluation has not yet been run; the full-set figures may differ slightly.
Usage
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="Blgn94/whisper-small-mn-v24")
result = asr("audio.mp3", generate_kwargs={"language": "mongolian", "task": "transcribe"})
print(result["text"])
Or with the model/processor directly:
import torch, librosa
from transformers import WhisperProcessor, WhisperForConditionalGeneration
processor = WhisperProcessor.from_pretrained("Blgn94/whisper-small-mn-v24", language="mongolian", task="transcribe")
model = WhisperForConditionalGeneration.from_pretrained("Blgn94/whisper-small-mn-v24")
audio, _ = librosa.load("audio.mp3", sr=16000)
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
ids = model.generate(inputs.input_features, language="mongolian", task="transcribe")
print(processor.batch_decode(ids, skip_special_tokens=True)[0])
Audio is expected at 16 kHz mono. Keep clips within Whisper's 30 s window.
Training data
Blgn94/mongolian-stt-dataset-v24
— a combined Mongolian read-speech and subtitle corpus (~77 h of audio, 16 kHz).
| Source | Train clips | Validation clips |
|---|---|---|
| Common Voice | 87,988 | 1,869 |
| FLEURS | 3,770 | 215 |
| MBSpeech | 3,670 | 176 |
| YouTube (manual subtitles) | 559 | 39 |
| Total | 95,987 | 2,299 |
Preprocessing: rows whose tokenized transcript exceeded 440 tokens (near Whisper's 448-token decoder limit) were dropped — 41 from train and 2 from validation — leaving 95,946 train / 2,297 validation clips.
Training procedure
Fine-tuned from openai/whisper-small (244M params).
| Hyperparameter | Value |
|---|---|
| Epochs | 3 |
| Effective batch size | 8 (per-device 2 × grad-accum 4) |
| Learning rate | 1e-5 |
| Warmup steps | 500 |
| Precision | fp16 |
| Total steps | 35,982 |
| Optimizer | AdamW (Transformers default) |
| Gradient checkpointing | enabled |
Hardware: single NVIDIA RTX 5060 (8 GB). Full run took ~6.5 h.
The model, task, and language tokens were set for Mongolian transcription
(language="mongolian", task="transcribe"); forced_decoder_ids was cleared
so decoding follows the fine-tuned distribution.
Limitations
- The training data is ~92% Common Voice (read speech), so accuracy is strongest on clear, read-style Mongolian and weaker on spontaneous, conversational, noisy, or telephony audio.
- Cyrillic Mongolian only.
- Inherits Whisper's tendency to hallucinate on silence or non-speech segments.
Base model & license
Derived from openai/whisper-small; released under Apache-2.0.