feat: swap Huggingface Transformers back-end to mlx native whisper.

calliope/app.py

@@ -6,9 +6,6 @@ import threading
 import time
 from typing import Any
 
-# Run offline — models are downloaded during setup, no need to hit HuggingFace on every launch.
-os.environ.setdefault("HF_HUB_OFFLINE", "1")
-
 import subprocess
 
 import rumps
@@ -36,7 +33,7 @@ class CalliopeApp(rumps.App):
         self.overlay = WaveformOverlay()
         self.recorder = Recorder(device=cfg.get("device"))
         self.transcriber = Transcriber(
-            model=cfg.get("model", "distil-whisper/distil-large-v3"),
+            model=cfg.get("model", "mlx-community/whisper-large-v3-turbo"),
             silence_threshold=cfg.get("silence_threshold", 0.005),
         )
         self.transcriber.context = cfg.get("context", "")
@@ -281,15 +278,11 @@ class CalliopeApp(rumps.App):
         log.info("Typing mode set to %s", mode)
 
     def _release_transcriber(self) -> None:
-        """Free the current Whisper model to reclaim GPU memory."""
+        """Free the current Whisper model to reclaim memory."""
         import gc
-        import torch
         if self.transcriber is not None:
-            self.transcriber._pipe = None
+            self.transcriber._loaded = False
-            self.transcriber._tokenizer = None
         gc.collect()
-        if torch.backends.mps.is_available():
-            torch.mps.empty_cache()
 
     def _on_toggle_click(self, sender) -> None:
         self._toggle_recording()

calliope/config.py

@@ -13,7 +13,7 @@ CONFIG_PATH = CONFIG_DIR / "config.yaml"
 
 DEFAULTS: dict[str, Any] = {
     "device": None,  # sounddevice index; None = system default
-    "model": "distil-whisper/distil-large-v3",
+    "model": "mlx-community/whisper-large-v3-turbo",
     "language": "auto",
     "hotkeys": {
         "ptt": "ctrl+shift",
@@ -50,11 +50,11 @@ LANGUAGES: dict[str, str] = {
 }
 
 MODELS: list[str] = [
-    "distil-whisper/distil-large-v3",
+    "mlx-community/whisper-large-v3-turbo",
-    "openai/whisper-large-v3",
+    "mlx-community/whisper-large-v3",
-    "openai/whisper-base",
+    "mlx-community/whisper-small",
-    "openai/whisper-small",
+    "mlx-community/whisper-medium",
-    "openai/whisper-medium",
+    "mlx-community/whisper-base",
 ]
 
 

calliope/transcriber.py

@@ -1,21 +1,19 @@
-"""Whisper transcription using transformers pipeline on MPS."""
+"""Whisper transcription using mlx-whisper on Apple Silicon."""
 
 import logging
 
 import numpy as np
-import torch
-from transformers import pipeline
 
 log = logging.getLogger(__name__)
 
+DEFAULT_MODEL = "mlx-community/whisper-large-v3-turbo"
+
 
 class Transcriber:
-    def __init__(self, model: str = "distil-whisper/distil-large-v3", silence_threshold: float = 0.005):
+    def __init__(self, model: str = DEFAULT_MODEL, silence_threshold: float = 0.005):
         self.model = model
-        self._pipe = None
+        self._loaded = False
-        self._tokenizer = None
         self._context: str = ""
-        self._cached_prompt_ids = None
         self.language: str = "auto"
         self.silence_threshold = silence_threshold
 
@@ -26,35 +24,24 @@ class Transcriber:
     @context.setter
     def context(self, value: str) -> None:
         self._context = value
-        self._cached_prompt_ids = None  # invalidate cache
 
     def load(self) -> None:
-        from transformers import AutoTokenizer
+        import mlx_whisper
 
-        device = "mps" if torch.backends.mps.is_available() else "cpu"
+        log.info("Loading model %s via mlx-whisper", self.model)
-        # Use float32 on MPS — float16 produces garbled output on Apple Silicon.
-        dtype = torch.float32 if device == "mps" else torch.float16
-        log.info("Loading model %s on %s (dtype=%s)", self.model, device, dtype)
         try:
-            self._pipe = pipeline(
+            mlx_whisper.transcribe(
-                "automatic-speech-recognition",
+                np.zeros(16_000, dtype=np.float32),
-                model=self.model,
+                path_or_hf_repo=self.model,
-                torch_dtype=dtype,
-                device=device,
-            )
-            self._tokenizer = AutoTokenizer.from_pretrained(self.model)
-            log.info("Model loaded, running warmup...")
-            self._pipe(
-                {"raw": np.zeros(16_000, dtype=np.float32), "sampling_rate": 16_000},
-                batch_size=1,
             )
+            self._loaded = True
             log.info("Model ready")
         except Exception:
             log.error("Failed to load model %s", self.model, exc_info=True)
             raise
 
     def transcribe(self, audio: np.ndarray) -> str:
-        if self._pipe is None:
+        if not self._loaded:
             self.load()
         if audio.size == 0:
             return ""
@@ -68,23 +55,15 @@ class Transcriber:
             log.debug("Audio too short or too quiet, skipping transcription")
             return ""
 
-        generate_kwargs = {}
+        import mlx_whisper
-        if self._context:
-            if self._cached_prompt_ids is None:
-                device = self._pipe.model.device
-                self._cached_prompt_ids = torch.tensor(self._tokenizer.get_prompt_ids(self._context), device=device)
-            generate_kwargs["prompt_ids"] = self._cached_prompt_ids
 
-        pipe_kwargs = {
+        kwargs = {
-            "batch_size": 1,
+            "path_or_hf_repo": self.model,
-            "generate_kwargs": generate_kwargs,
         }
+        if self._context:
+            kwargs["initial_prompt"] = self._context
         if self.language != "auto":
-            pipe_kwargs["generate_kwargs"]["language"] = self.language
+            kwargs["language"] = self.language
-            pipe_kwargs["generate_kwargs"]["task"] = "transcribe"
 
-        result = self._pipe(
+        result = mlx_whisper.transcribe(audio, **kwargs)
-            {"raw": audio, "sampling_rate": 16_000},
-            **pipe_kwargs,
-        )
         return result["text"].strip()

pyproject.toml

@@ -11,9 +11,7 @@ dependencies = [
     "rumps>=0.4.0",
     "sounddevice>=0.4.6",
     "numpy>=1.24.0",
-    "torch>=2.0.0",
+    "mlx-whisper>=0.4.0",
-    "transformers>=4.36.0",
-    "accelerate>=0.25.0",
     "pynput>=1.7.6",
     "pyobjc-framework-Quartz>=9.0",
     "pyobjc-framework-Cocoa>=9.0",