DILHTWD/whisper-large-v3-hsb

Model Description

This model was fine-tuned on over 24 hours of transcribed upper sorbian speech to aid future research, conservation and revitalisation of the language.

Training Data

• Source: Stiftung für das sorbische Volk / Załožba za serbski lud (https://stiftung.sorben.com/)
• Volume: 1493 Minutes, 10% Validation Set, 10% Test Set

Training Details

• Hyperparameters:
  • Batch size: 64
  • Learning rate: 3e-6, linear decay
• Optimizer: AdamW
• Warmup: 1000 steps
• Additional Techniques: BF16 training, initial 15 layers frozen

Performance

Metrics

• Word Error Rate: 5.7

Usage

Example Code

To use the model, follow this example code:

import torch
import torchaudio
from transformers import WhisperProcessor, WhisperForConditionalGeneration

# Load the model and processor
model_name = "DILHTWD/whisper-large-v3-hsb"
processor_name = "openai/whisper-large-v3"
processor = WhisperProcessor.from_pretrained(processor_name)
model = WhisperForConditionalGeneration.from_pretrained(model_name)

# Load and preprocess the audio
audio, sample_rate = torchaudio.load("test.mp3")
if sample_rate != 16000:
    audio = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)(audio)
input_features = processor(audio.squeeze().numpy(), sampling_rate=16000, return_tensors="pt").input_features

# Generate transcription
with torch.no_grad():
    predicted_ids = model.generate(input_features)
    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

# Print the transcription
print("Transcription:", transcription)

Model Details

• Model Name: DILHTWD/whisper-large-v3-hsb
• Publisher: Data Intelligence Lab, Hochschule für Technik und Wirtschaft Dresden
• Model Version: 1.0.0
• Model Date: 2024-11-11
• License: AGPL-3.0
• Architecture: Whisper Large v3
• Task: Automatic Speech Recognition