import os import spaces import gradio as gr from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor import torch import numpy as np import librosa MODEL_REPO = "aee4/wav2vec2-xlsr-akan-v3" model = Wav2Vec2ForCTC.from_pretrained(MODEL_REPO) processor = Wav2Vec2Processor.from_pretrained(MODEL_REPO) model.eval() @spaces.GPU def transcribe(audio): if audio is None: return "Upload or record audio first." sr, data = audio if data.dtype != np.float32: data = data.astype(np.float32) / np.iinfo(data.dtype).max if data.ndim > 1: data = data.mean(axis=1) # resample to 16kHz if needed, model was trained at 16kHz if sr != 16000: data = librosa.resample(data, orig_sr=sr, target_sr=16000) sr = 16000 model.to("cuda") input_values = processor(data, sampling_rate=sr, return_tensors="pt").input_values.to("cuda") with torch.no_grad(): logits = model(input_values).logits pred_ids = torch.argmax(logits, dim=-1) return processor.batch_decode(pred_ids)[0] demo = gr.Interface( fn=transcribe, inputs=gr.Audio(sources=["microphone", "upload"], type="numpy"), outputs=gr.Textbox(label="Transcription"), title="Akan ASR — Wav2Vec2 XLS-R", description="Fine-tuned on Paywinful/akan_asr_dataset. WER 0.319 at step 5500." ) demo.launch()