Switch to faster-whisper

README.md

@@ -6,18 +6,18 @@ Python X11 transcription daemon that records audio, runs Whisper, logs the trans
 
 - X11 (not Wayland)
 - `ffmpeg`
-- `whisper` (OpenAI Whisper CLI)
+- `faster-whisper`
 - `xclip`
 - `xdotool`
 - Tray icon deps: `libappindicator3` and `gtk3` (required by `systray`)
-- Python deps: `pystray`, `pillow`, `python-xlib`, `ollama`, `openai-whisper`
+- Python deps: `pystray`, `pillow`, `python-xlib`, `ollama`, `faster-whisper`
 
 ## Python Daemon
 
 Install Python deps:
 
 ```bash
-pip install -r src/requirements.txt
+pip install -r requirements.txt
 ```
 
 Run:

requirements.txt

@@ -1,4 +1,4 @@
-openai-whisper
+faster-whisper
 ollama
 pystray
 pillow

src/leld.py

@@ -11,7 +11,7 @@ from pathlib import Path
 
 from config import Config, load, redacted_dict
 from recorder import start_recording, stop_recording
-from stt import WhisperSTT
+from stt import FasterWhisperSTT, STTConfig
 from aiprocess import AIConfig, build_processor
 from inject import inject
 from x11_hotkey import listen
@@ -34,7 +34,14 @@ class Daemon:
         self.proc = None
         self.record = None
         self.timer = None
-        self.stt = WhisperSTT(cfg.whisper_model, cfg.whisper_lang, cfg.whisper_device)
+        self.stt = FasterWhisperSTT(
+            STTConfig(
+                model=cfg.whisper_model,
+                language=cfg.whisper_lang,
+                device=cfg.whisper_device,
+                vad_filter=True,
+            )
+        )
         self.ai = None
         if cfg.ai_enabled:
             self.ai = build_processor(

src/stt.py

@@ -1,25 +1,48 @@
-import os
-import whisper
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from faster_whisper import WhisperModel
 
 
-def _force_cpu():
-    os.environ.setdefault("CUDA_VISIBLE_DEVICES", "")
+@dataclass
+class STTConfig:
+    model: str
+    language: str | None
+    device: str
+    vad_filter: bool
 
 
-class WhisperSTT:
-    def __init__(self, model: str, language: str | None = None, device: str = "cpu"):
-        self.model_name = model
-        self.language = language
-        self.device = (device or "cpu").lower()
-        self._model = None
+def _compute_type(device: str) -> str:
+    dev = (device or "cpu").lower()
+    if dev == "cuda":
+        return "float16"
+    return "int8"
+
+
+class FasterWhisperSTT:
+    def __init__(self, cfg: STTConfig):
+        self.cfg = cfg
+        self._model: WhisperModel | None = None
 
     def _load(self):
         if self._model is None:
-            if self.device == "cpu":
-                _force_cpu()
-            self._model = whisper.load_model(self.model_name, device=self.device)
+            self._model = WhisperModel(
+                self.cfg.model,
+                device=self.cfg.device or "cpu",
+                compute_type=_compute_type(self.cfg.device),
+            )
 
     def transcribe(self, wav_path: str) -> str:
         self._load()
-        result = self._model.transcribe(wav_path, language=self.language)
-        return (result.get("text") or "").strip()
+        segments, _info = self._model.transcribe(
+            wav_path,
+            language=self.cfg.language,
+            vad_filter=self.cfg.vad_filter,
+        )
+        parts = []
+        for seg in segments:
+            text = (seg.text or "").strip()
+            if text:
+                parts.append(text)
+        return " ".join(parts).strip()