utachicodes/dikkte_asr

Wolof speech recognition. # Dikkte ASR Wolof speech-to-text. Fine-tunes Whisper small with LoRA, merges the adapter, and gives you a standard `transformers` model you can load in two lines. Wolof is spoken by 10M+ people across Senegal, Gambia, and Mauritania — but barely has any open ASR tools. This project fixes that. ## Install ```bash git clone github.com cd dikkte_asr pip install -r requirements.txt ``` ## Usage ### From HuggingFace (easiest) ```bash pip install transformers torch torchaudio ``` ```python from transformers import pipeline pipe = pipeline("automatic-speech-recognition", model="utachicodes/dikkte-wolof-asr") print(pipe("audio.wav")["text"]) ``` ### Load the model directly ```python from transformers import WhisperForConditionalGeneration, WhisperProcessor import torchaudio, torch processor = WhisperProcessor.from_pretrained("utachicodes/dikkte-wolof-asr") model = WhisperForConditionalGeneration.from_pretrained("utachicodes/dikkte-wolof-asr") model.eval() waveform, sr = torchaudio.load("audio.wav") if sr != 16000: waveform = torchaudio.transforms.Resample(sr, 16000)(waveform) inputs = processor(waveform.squeeze().numpy(), sampling_rate=16000, return_tensors="pt") with torch.no_grad(): ids = model.generate(input_features=inputs.input_features) print(processor.batch_decode(ids, skip_special_tokens=True)[0]) ``` ### Web UI (mic input) ```bash python wolof_stt.py # opens at 127.0.0.1 ``` Record from your mic, hit transcribe. Long audio gets chunked into 30s segments. ### Use the raw LoRA adapter If you want to apply the adapter yourself instead of using the merged model: ```python from transformers import WhisperForConditionalGeneration, WhisperProcessor from peft import PeftModel base = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small") model = PeftModel.from_pretrained(base, "./wolof-whisper-small-lora") model = model.merge_and_unload() ``` ## Training | | | |---|---| | Base model | `o …