Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 11·5 min read
BaltiVoice: A Speech Corpus and Fine-tuned Whisper ASR System for the Balti Language
TL;DR
The Balti language lacks sufficient speech data for effective automatic speech recognition (ASR) systems. A speech corpus named BaltiVoice was created, along with a fine-tuned Whisper ASR model specifically for Balti.
✦ Why It Matters
Engineers can leverage the BaltiVoice corpus and Whisper ASR model to create inclusive language technologies.
Key Takeaways
- Utilize the BaltiVoice corpus for ASR research in low-resource languages.
- Experiment with fine-tuning Whisper models for improved transcription accuracy.
- Access the dataset and models on HuggingFace for practical applications.
- Consider model capacity when developing ASR systems for underrepresented languages.
Full Summary
BaltiVoice introduces a 16.8-hour speech corpus for the Balti language, which previously lacked publicly available ASR resources. This corpus consists of 10,060 validated utterances in the native Nastaliq script, sourced from Mozilla Common Voice recordings.
The researchers fine-tuned the OpenAI Whisper-small model, resulting in a Word Error Rate (WER) of 26.74% and a Character Error Rate (CER) of 8.67% on a validation set of 538 utterances. This is a substantial improvement from a zero-shot baseline of 159.19% WER and 152.52% CER.
Additionally, a Whisper-base model fine-tuned on the same data achieved a WER of 44.54% and a CER of 15.61%, indicating that model capacity is crucial in low-resource settings. The dataset, along with the fine-tuned model and a live transcription demo, is publicly accessible on HuggingFace.