feat(workers): whisper loader with CUDA detect + CPU fallback

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

workers/void_workers/model.py

@@ -0,0 +1,34 @@
+import os
+from .log import log
+
+_whisper_model = None
+
+
+def cuda_available():
+    try:
+        import ctranslate2
+        return ctranslate2.get_cuda_device_count() > 0
+    except Exception as e:
+        log.info("ctranslate2_cuda_probe_failed", err=str(e))
+        return False
+
+
+def whisper_model():
+    global _whisper_model
+    if _whisper_model is None:
+        from faster_whisper import WhisperModel
+        name = os.environ.get("WHISPER_MODEL", "small.en")
+        cache = os.environ.get("WHISPER_CACHE", "/var/lib/void/whisper-models")
+        device = "cuda" if cuda_available() else "cpu"
+        compute_type = "float16" if device == "cuda" else "int8"
+        log.info("whisper_loading", model=name, device=device,
+                 compute_type=compute_type, cache=cache)
+        _whisper_model = WhisperModel(
+            name, device=device, compute_type=compute_type, download_root=cache
+        )
+    return _whisper_model
+
+
+def whisper_transcribe(audio_path):
+    segments, _info = whisper_model().transcribe(audio_path, vad_filter=True)
+    return "\n".join(s.text.strip() for s in segments).strip()