Spaces:
Sleeping
Sleeping
| #!/usr/bin/env python3 | |
| # ------------------------------------------------------------ | |
| # Whisper Small LS-100 Demo (local or Hugging Face Spaces) | |
| # Author: Murali Chandran (codeninja3d) | |
| # ------------------------------------------------------------ | |
| import torch | |
| import torchaudio | |
| import gradio as gr | |
| from transformers import WhisperProcessor, WhisperForConditionalGeneration | |
| # ------------------------------- | |
| # Config | |
| # ------------------------------- | |
| MODEL_ID = "codeninja3d/whisper-small-ls100" | |
| DEVICE = "cuda" if torch.cuda.is_available() else "cpu" | |
| # print(f"✅ Loading model from: {MODEL_ID}") | |
| processor = WhisperProcessor.from_pretrained(MODEL_ID) | |
| model = WhisperForConditionalGeneration.from_pretrained(MODEL_ID).to(DEVICE).eval() | |
| # print(f"✅ #Params: {sum(p.numel() for p in model.parameters()) / 1e6:.2f}M") | |
| # print("✅ Sample rate:", processor.feature_extractor.sampling_rate) | |
| # ------------------------------- | |
| # Transcribe function | |
| # ------------------------------- | |
| def transcribe(audio, language, task): | |
| if audio is None: | |
| return "⚠️ Please record or upload an audio clip." | |
| # Handle tuple (mic) or file path (upload) | |
| if isinstance(audio, tuple): | |
| sr, waveform_np = audio | |
| waveform = torch.tensor(waveform_np, dtype=torch.float32) | |
| if waveform.ndim > 1: # [samples, channels] → [1, samples] | |
| waveform = waveform.mean(dim=-1, keepdim=True).T | |
| else: | |
| waveform, sr = torchaudio.load(audio) | |
| # Ensure mono and correct shape [1, samples] | |
| if waveform.ndim > 1: | |
| waveform = waveform.mean(dim=0, keepdim=True) | |
| # Resample to 16 kHz | |
| if sr != 16000: | |
| waveform = torchaudio.functional.resample(waveform, sr, 16000) | |
| sr = 16000 | |
| # ✅ Normalize amplitude to [-1, 1] | |
| waveform = waveform / waveform.abs().max() | |
| # ✅ Force 1D numpy array (no 2D padding issues) | |
| waveform_np = waveform.squeeze().numpy().astype("float32") | |
| # Preprocess | |
| inputs = processor.feature_extractor( | |
| waveform_np, | |
| sampling_rate=sr, | |
| return_tensors="pt" | |
| ).to(DEVICE) | |
| # print(f"[DEBUG] waveform shape: {waveform.shape}, dtype: {waveform.dtype}, sr={sr}") | |
| # print(f"[DEBUG] min={waveform.min().item():.3f}, max={waveform.max().item():.3f}, mean={waveform.mean().item():.3f}") | |
| # Decoder setup | |
| forced_decoder_ids = processor.get_decoder_prompt_ids(language="en", task="transcribe") | |
| # print("[DEBUG] forced_decoder_ids:", forced_decoder_ids) | |
| with torch.no_grad(): | |
| pred_ids = model.generate( | |
| **inputs, | |
| forced_decoder_ids=forced_decoder_ids, | |
| max_new_tokens=256, | |
| num_beams=5, | |
| no_repeat_ngram_size=3, | |
| repetition_penalty=1.2, | |
| early_stopping=True | |
| ) | |
| text = processor.tokenizer.batch_decode(pred_ids, skip_special_tokens=True)[0] | |
| return text.strip() | |
| # ------------------------------- | |
| # Gradio Interface | |
| # ------------------------------- | |
| iface = gr.Interface( | |
| fn=transcribe, | |
| inputs=[ | |
| gr.Audio(type="numpy", label="🎙️ Record or Upload Audio"), | |
| ], | |
| outputs=gr.Textbox(label="📄 Transcription"), | |
| title="Whisper Small LS-100 (Fine-tuned)", | |
| description="Upload or record audio to test Murali's fine-tuned Whisper model on LibriSpeech-100.", | |
| theme="soft", | |
| allow_flagging="auto" | |
| ) | |
| iface.launch(server_name="0.0.0.0", server_port=7860) | |