Whisper Sinhala Proto

--- language: - si license: apache-2.0 library_name: transformers pipeline_tag: automatic-speech-recognition base_model: openai/whisper-small datasets: - large-sinhala-asr-dataset tags: - whisper - speech-recognition - sinhala - automatic-speech-recognition - prototype --- # Whisper Sinhala Prototype `whisper-sinhala-proto` is an experimental Sinhala speech-recognition model based on [OpenAI Whisper Small](https://huggingface.co/openai/whisper-small). The checkpoint is intended for early testing, benchmarking, and comparison during Sinhala ASR development. ## Repository status This model appears to share the same base architecture, parameter count, task, and training-dataset tag as `kaan84/whisper-small-sinhala-proto`. If both repositories contain equivalent weights, consider keeping one as the canonical prototype and marking the other as archived or superseded to reduce user confusion. ## Model details | Item | Value | | --- | --- | | Task | Automatic speech recognition | | Language | Sinhala (`si`) | | Base model | `openai/whisper-small` | | Model class | `AutoModelForSpeechSeq2Seq` | | Parameters | Approximately 241.7M | | Training dataset tag | `large-sinhala-asr-dataset` | | Status | Experimental prototype | ## Usage ```python from transformers import pipeline transcriber = pipeline( "automatic-speech-recognition", model="kaan84/whisper-sinhala-proto", ) result = transcriber( "sample.wav", generate_kwargs={"language": "si", "task": "transcribe"}, ) print(result["text"]) ``` ## Recommended use - Sinhala ASR experiments - Baseline and checkpoint comparison - Error analysis across accents and recording conditions - Further fine-tuning on appropriately licensed Sinhala speech data This prototype is not documented or evaluated sufficiently for high-impact production use. ## Training and evaluation information needed To complete this model card, add: - Dataset source, license, duration, speakers, and split sizes - Audio resampling and transcript-normalization rules - Training hyperparameters and hardware - Validation and test WER/CER - Comparison with the base `openai/whisper-small` model - Known failure cases and subgroup analysis ## Limitations The model may hallucinate, omit, or incorrectly transcribe speech. Performance can vary with audio quality, noise, code-switching, accent, vocabulary, and speaker characteristics. Human revie...

Source context: 13 downloads · 0 likes · Pipeline automatic-speech-recognition · Library transformers · Repo kaan84/whisper-sinhala-proto