diff --git a/backend/whisper_admin.py b/backend/whisper_admin.py
index 19413e7..3c60d20 100644
--- a/backend/whisper_admin.py
+++ b/backend/whisper_admin.py
@@ -3,6 +3,9 @@ from __future__ import annotations
 import importlib
 import importlib.util
 import os
+import shutil
+import subprocess
+import tempfile
 import threading
 from pathlib import Path
 from typing import Any, Dict, Optional
@@ -10,6 +13,20 @@ from typing import Any, Dict, Optional
 
 DEFAULT_WHISPER_MODEL = "base"
 _WHISPER_DOWNLOAD_LOCK = threading.Lock()
+_WHISPER_MODEL_LOCK = threading.Lock()
+_WHISPER_MODELS: Dict[tuple[str, str], Any] = {}
+_AUDIO_SUFFIXES = {
+    "audio/aac": ".aac",
+    "audio/flac": ".flac",
+    "audio/mp4": ".m4a",
+    "audio/mpeg": ".mp3",
+    "audio/mpga": ".mp3",
+    "audio/ogg": ".ogg",
+    "audio/wav": ".wav",
+    "audio/wave": ".wav",
+    "audio/webm": ".webm",
+    "audio/x-wav": ".wav",
+}
 
 
 def _default_download_root() -> Path:
@@ -24,6 +41,13 @@ def _load_whisper_module():
         return None
 
 
+def _load_torch_module():
+    try:
+        return importlib.import_module("torch")
+    except Exception:
+        return None
+
+
 def whisper_runtime_error() -> Optional[str]:
     if importlib.util.find_spec("whisper") is None:
         return (
@@ -123,3 +147,126 @@ def ensure_whisper_model_downloaded(model_name: str = DEFAULT_WHISPER_MODEL) ->
         "path": str(target),
         "error": "",
     }
+
+
+def _resolve_whisper_device() -> str:
+    try:
+        torch_mod = _load_torch_module()
+        if torch_mod is not None and getattr(torch_mod.cuda, "is_available", lambda: False)():
+            return "cuda"
+
+        backends = getattr(torch_mod, "backends", None)
+        mps_backend = getattr(backends, "mps", None)
+        if mps_backend is not None and getattr(mps_backend, "is_available", lambda: False)():
+            return "mps"
+    except Exception:
+        pass
+
+    return "cpu"
+
+
+def _resolve_ffmpeg_binary() -> Optional[str]:
+    candidate = os.getenv("HEIMGEIST_FFMPEG_PATH") or shutil.which("ffmpeg") or "/usr/bin/ffmpeg"
+    if not candidate:
+        return None
+    if Path(candidate).exists() or shutil.which(candidate):
+        return str(candidate)
+    return None
+
+
+def _audio_suffix_for_mime_type(mime_type: str) -> str:
+    base_mime = str(mime_type or "").split(";", 1)[0].strip().lower()
+    return _AUDIO_SUFFIXES.get(base_mime, ".webm")
+
+
+def _load_transcription_model(model_name: str = DEFAULT_WHISPER_MODEL) -> tuple[Any, str]:
+    error = whisper_runtime_error()
+    if error:
+        raise RuntimeError(error)
+
+    whisper_mod = _load_whisper_module()
+    if whisper_mod is None:
+        raise RuntimeError("Failed to import the Whisper runtime.")
+
+    ensure_whisper_model_downloaded(model_name)
+    device = _resolve_whisper_device()
+    cache_key = (model_name, device)
+
+    with _WHISPER_MODEL_LOCK:
+        model = _WHISPER_MODELS.get(cache_key)
+        if model is None:
+            try:
+                model = whisper_mod.load_model(model_name, device=device)
+            except TypeError:
+                model = whisper_mod.load_model(model_name)
+            _WHISPER_MODELS[cache_key] = model
+
+    return model, device
+
+
+def _convert_audio_to_wav(input_path: Path, output_path: Path) -> None:
+    ffmpeg_bin = _resolve_ffmpeg_binary()
+    if not ffmpeg_bin:
+        raise RuntimeError(
+            "Audio transcription requires ffmpeg. Heimgeist could not resolve it from HEIMGEIST_FFMPEG_PATH or PATH."
+        )
+
+    process = subprocess.run(
+        [
+            ffmpeg_bin,
+            "-y",
+            "-hide_banner",
+            "-loglevel",
+            "error",
+            "-i",
+            str(input_path),
+            "-ac",
+            "1",
+            "-ar",
+            "16000",
+            "-f",
+            "wav",
+            str(output_path),
+        ],
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        text=True,
+    )
+    if process.returncode != 0:
+        detail = (process.stderr or process.stdout or "ffmpeg audio conversion failed").strip()
+        raise RuntimeError(detail)
+
+
+def transcribe_audio_bytes(
+    audio_bytes: bytes,
+    mime_type: str,
+    model_name: str = DEFAULT_WHISPER_MODEL,
+) -> Dict[str, Any]:
+    if not audio_bytes:
+        raise RuntimeError("Recorded audio was empty.")
+
+    model, device = _load_transcription_model(model_name)
+    input_path = Path(tempfile.mkstemp(suffix=_audio_suffix_for_mime_type(mime_type))[1])
+    wav_path = Path(tempfile.mkstemp(suffix=".wav")[1])
+
+    try:
+        input_path.write_bytes(audio_bytes)
+        source_path = input_path
+        if input_path.suffix.lower() != ".wav":
+            _convert_audio_to_wav(input_path, wav_path)
+            source_path = wav_path
+
+        result = model.transcribe(str(source_path), task="transcribe", fp16=device == "cuda")
+        return {
+            "model": model_name,
+            "device": device,
+            "language": str(result.get("language") or "").strip(),
+            "text": str(result.get("text") or "").strip(),
+        }
+    finally:
+        for path in (input_path, wav_path):
+            try:
+                if path.exists():
+                    path.unlink()
+            except Exception:
+                pass