Korla/Wav2Vec2BertForCTC-hsb

Model Description

Wav2Vec2BertForCTC-hsb is a fine-tuned Wav2Vec2 model with a BERT-style character classification head, adapted for Upper Sorbian automatic speech recognition (ASR). This model has been fine-tuned for CTC (Connectionist Temporal Classification) loss and is capable of transcribing audio in the Upper Sorbian language.

Usage

This model can be used for speech-to-text tasks on Upper Sorbian audio.

An optional 5-gram language model (5gram.bin) is provided for decoding with an external LM scorer. This n-gram model was trained on a corpus of Upper Sorbian Holy Masses, which can help improve decoding accuracy for religious or formal speech domains.

Training Data

The model was fine-tuned on a dataset provided by the Foundation for the Sorbian People, which consists of high-quality recordings and transcripts in Upper Sorbian. The dataset includes diverse speakers and speech conditions, ensuring a robust acoustic model.

Language Model

• Name: 5gram.bin
• Type: 5-gram character-level KenLM language model
• Domain: Upper Sorbian religious speech (Holy Masses)
• Usage: For decoding with tools such as CTCDecoder.

Limitations

• The model's accuracy may degrade on informal or highly dialectal speech not represented in the training data.
• The language model is domain-specific (religious speech) and may bias decoding toward that context.
• The model supports only Upper Sorbian, not Lower Sorbian or other Slavic languages.

How to Use

For normal use (without LM) you can load the model into a pipeline.

To use the 5-gram language model for decoding, use the pyctcdecode library.

Citation

Please cite as:

@misc{korla_wav2vec2bertforctc_hsb,
  author       = {Karl Baier},
  title        = {Wav2Vec2BertForCTC-hsb},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/Korla/Wav2Vec2BertForCTC-hsb}},
}