agirdi

sachmatkris/agirdi is a Lithuanian automatic speech recognition (ASR) model fine-tuned from openai/whisper-large-v3-turbo.

The model was fine-tuned specifically for Lithuanian speech recognition using the read (read) and spontaneous (spon) portions of the LIEPA-3 corpus.

Model Details

• Base model: openai/whisper-large-v3-turbo
• Language: Lithuanian (lt)
• Task: Automatic Speech Recognition (ASR)
• Architecture: Whisper
• License: MIT

Training Data

agirdi was fine-tuned on the LIEPA-3 (Didysis lietuvių kalbos garsynas) Lithuanian speech corpus.

Only the following LIEPA-3 subsets were used for fine-tuning:

• read — read Lithuanian speech
• spon — spontaneous Lithuanian speech

The remaining LIEPA-3 subsets were not used for training.

LIEPA-3 is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Evaluation

The model is evaluated using Word Error Rate (WER).

Dataset	WER ↓
LIEPA-3 test set	3.05%
FLEURS Lithuanian	10.50%

The LIEPA-3 test set measures performance on speech from the same corpus used for fine-tuning, while FLEURS Lithuanian provides an external evaluation set for measuring generalization beyond the training corpus.

Usage

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

model_id = "sachmatkris/agirdi"

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
    use_safetensors=True,
)

model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

result = pipe("audio.wav")

print(result["text"])

Limitations

agirdi is specialized for Lithuanian speech recognition. Performance may vary depending on recording quality, background noise, speaker characteristics, dialect, domain-specific terminology, and other acoustic conditions.

The model may also inherit limitations and biases present in the original Whisper model and the LIEPA-3 training data.

Acknowledgements

agirdi was fine-tuned from openai/whisper-large-v3-turbo using data from the LIEPA-3 Lithuanian speech corpus.

Please refer to the LIEPA-3 dataset documentation for its license, attribution requirements, and further information about the corpus.