Initial public CPU Whisper small STT space

README.md

@@ -1,12 +1,18 @@
 ---
 title: Whisper Small STT CPU
-emoji: 🐠
-colorFrom: purple
-colorTo: blue
+emoji: 🎙️
+colorFrom: blue
+colorTo: indigo
 sdk: gradio
-sdk_version: 6.12.0
+sdk_version: 5.25.1
 app_file: app.py
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Whisper Small STT (Free CPU)
+
+Public speech-to-text Space powered by `openai/whisper-small` on free CPU hardware.
+
+- Upload audio or record with your microphone
+- Choose `transcribe` (same language) or `translate` (to English)
+- Includes a public API endpoint: `/transcribe`

app.py

@@ -0,0 +1,86 @@
+import numpy as np
+import torch
+import gradio as gr
+from transformers import WhisperForConditionalGeneration, WhisperProcessor
+
+MODEL_ID = "openai/whisper-small"
+
+processor = WhisperProcessor.from_pretrained(MODEL_ID)
+model = WhisperForConditionalGeneration.from_pretrained(MODEL_ID)
+model.eval()
+
+
+def _to_float32_audio(audio: tuple[int, np.ndarray]) -> tuple[int, np.ndarray]:
+    sample_rate, data = audio
+    if data.ndim > 1:
+        data = data.mean(axis=1)
+
+    if np.issubdtype(data.dtype, np.integer):
+        max_int = np.iinfo(data.dtype).max
+        data = data.astype(np.float32) / float(max_int)
+    else:
+        data = data.astype(np.float32)
+        peak = np.max(np.abs(data)) if data.size else 0.0
+        if peak > 1.0:
+            data = data / peak
+
+    return sample_rate, data
+
+
+def transcribe_audio(audio, task):
+    if audio is None:
+        return "Please upload or record audio first."
+
+    sample_rate, data = _to_float32_audio(audio)
+
+    inputs = processor(
+        data,
+        sampling_rate=sample_rate,
+        return_tensors="pt",
+    )
+
+    forced_decoder_ids = processor.get_decoder_prompt_ids(task=task)
+
+    with torch.inference_mode():
+        predicted_ids = model.generate(
+            inputs.input_features,
+            forced_decoder_ids=forced_decoder_ids,
+        )
+
+    text = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0].strip()
+    return text or "(No speech detected)"
+
+
+with gr.Blocks(title="Whisper Small STT CPU") as demo:
+    gr.Markdown(
+        """
+# Whisper Small STT (Free CPU)
+Public speech-to-text using `openai/whisper-small`.
+
+- `transcribe`: keep original language
+- `translate`: translate speech to English
+"""
+    )
+
+    audio_input = gr.Audio(
+        label="Audio Input",
+        type="numpy",
+        sources=["upload", "microphone"],
+    )
+    task_input = gr.Dropdown(
+        choices=["transcribe", "translate"],
+        value="transcribe",
+        label="Task",
+    )
+    run_btn = gr.Button("Convert Speech to Text", variant="primary")
+    text_output = gr.Textbox(label="Transcript", lines=12)
+
+    run_btn.click(
+        fn=transcribe_audio,
+        inputs=[audio_input, task_input],
+        outputs=[text_output],
+        api_name="transcribe",
+    )
+
+if __name__ == "__main__":
+    demo.queue(default_concurrency_limit=1).launch()

requirements.txt

@@ -0,0 +1,4 @@
+gradio>=5.25.0
+transformers>=4.46.0
+torch
+numpy