Use faster-whisper

Dockerfile

@@ -1,18 +1,15 @@
-FROM python:3.10
+FROM ghcr.io/astral-sh/uv:python3.12-bookworm-slim AS base
+
+FROM base AS model-downloader
 
-# Pre-download Whisper model using ADD
-ARG WHISPER_MODEL_URL=https://openaipublic.azureedge.net/main/whisper/models/345ae4da62f9b3d59415adc60127b97c714f32e89e936602e85993674d08dcb1/medium.pt
 ARG WHISPER_MODEL=medium
-RUN mkdir -p /root/.cache/whisper
-ADD ${WHISPER_MODEL_URL} /root/.cache/whisper/${WHISPER_MODEL}.pt
+RUN uv pip install --system faster-whisper
+RUN python -c "from faster_whisper import WhisperModel; model = WhisperModel('${WHISPER_MODEL}', device='cpu')"
 
-# Remove apt auto-clean hook to preserve cache
-RUN rm -f /etc/apt/apt.conf.d/docker-clean
+FROM base
 
-# Install system dependencies with cache mount
-RUN --mount=type=cache,target=/var/cache/apt \
-    --mount=type=cache,target=/var/lib/apt/lists \
-    apt update && apt install -y --no-install-recommends ffmpeg
+# Copy model cache from first stage
+COPY --from=model-downloader /root/.cache /root/.cache
 
 # Copy uv binary
 COPY --from=ghcr.io/astral-sh/uv:latest /uv /bin/uv

app.py

@@ -5,9 +5,8 @@
 
 import dash
 import dash_bootstrap_components as dbc
-import torch
-import whisper
 from dash import Input, Output, State, dcc, html
+from faster_whisper import WhisperModel
 
 UPLOAD_FOLDER = tempfile.mkdtemp(dir=os.getcwd())
 
@@ -65,12 +64,27 @@ def write_srt(transcript: Iterator[dict], file: TextIO):
 
 
 def translate_transcribe_file(file_path):
-    try:
-        model = whisper.load_model("medium")
-    except torch.OutOfMemoryError:  # fallback
-        model = whisper.load_model("tiny", device="cpu")
-    translation = model.transcribe(file_path, language="no", task="translate")
-    transcription = model.transcribe(file_path, language="no")
+    model = WhisperModel("medium")
+
+    # Get translation segments
+    translation_segments, _ = model.transcribe(
+        file_path, language="no", task="translate"
+    )
+    translation = {
+        "segments": [
+            {"start": seg.start, "end": seg.end, "text": seg.text}
+            for seg in translation_segments
+        ]
+    }
+
+    # Get transcription segments
+    transcription_segments, _ = model.transcribe(file_path, language="no")
+    transcription = {
+        "segments": [
+            {"start": seg.start, "end": seg.end, "text": seg.text}
+            for seg in transcription_segments
+        ]
+    }
 
     # Create temporary files for translation and transcription
     with tempfile.NamedTemporaryFile(

pyproject.toml

@@ -10,5 +10,6 @@ requires-python = ">=3.8"
 dependencies = [
     "dash==2.14.0",
     "dash-bootstrap-components>=1.5.0",
-    "openai-whisper>=20250625",
+    "faster-whisper>=1.1.0",
 ]
+