Adding translation support for any language, adding mlx backend for s…

.env.example

@@ -1,12 +1,30 @@
 DATABASE_URL=postgresql://transcriber:transcriber@localhost:5433/transcriber
 REDIS_URL=redis://localhost:6380/0
 LLM_PROVIDER=ollama
+# To use a local openai compatible server set the base url here, for example:
+#OPENROUTER_BASE_URL=http://127.0.0.1:8000/v1/
+# An API key may not be needed when using your own server
 OPENROUTER_API_KEY=your_key_here
 OPENROUTER_MODEL=anthropic/claude-sonnet-4
+
 OLLAMA_BASE_URL=http://localhost:11434
 OLLAMA_MODEL=qwen3:8b
+
+# Whisper CLI path is not required if using the MLX backend.
 WHISPER_CLI_PATH=/path/to/whisper-cli
 WHISPER_MODEL_PATH=/path/to/kb_whisper_ggml_medium.bin
 WHISPER_SMALL_MODEL_PATH=/path/to/kb_whisper_ggml_small.bin
 STORAGE_PATH=./storage
+
+# Hugging Face authentication token, required for some models and features. Get it from https://huggingface.co/settings/tokens
 HF_AUTH_TOKEN=hf_your_token_here
+
+# Supported backends: mlx, torch
+#BACKEND=mlx
+
+# When using a mixture of Docker and host services, it can be helpful to uncomment these two settings:
+# LISTEN_HOST=0.0.0.0              # Allow requests from outside the docker network
+# PROXY_HOST=host.docker.internal  # Allow Docker containers to access services running on the host machine
+
+# If a non-standard API port is needed, uncomment and set the value
+#API_PORT=8082

README.md

@@ -12,7 +12,7 @@ AI-powered local meeting transcription with automatic speaker identification. Up
 1. **Upload, record, or go live** through the web UI
 2. **Audio extraction** - FFmpeg converts to 16kHz mono WAV
 3. **Transcription** - whisper.cpp with KB-LAB Swedish models (Metal GPU accelerated)
-4. **Speaker diarization** - pyannote.audio 3.1 separates speakers
+4. **Speaker diarization** - pyannote.audio 3.1 on torch, or MLX sortformer on Apple Silicon
 5. **Intro analysis** - LLM iteratively reads the transcript to detect introductions and count speakers
 6. **Speaker identification** - Names matched to voices using LLM reasoning + SpeechBrain voice embeddings
 7. **Results** - Color-coded transcript synced with audio playback, editable segments, AI-powered actions, export to 7 formats
@@ -50,7 +50,7 @@ The instructions below are for **macOS with Apple Silicon**. For other platforms
 - **FFmpeg** (`brew install ffmpeg`)
 - **whisper.cpp** compiled with Metal support
 - **Ollama** with a model like `qwen3:8b` (recommended), or an OpenRouter API key
-- **Hugging Face token** with access to `pyannote/speaker-diarization-3.1`
+- **Hugging Face token** with access to `pyannote/speaker-diarization-3.1` when using the `torch` diarization backend
 
 ## Quick install
 
@@ -109,6 +109,32 @@ This starts:
 - PostgreSQL on port **5433**
 - Redis on port **6380**
 
+### Docker networking gotchas (important)
+
+When mixing host-native services with Docker containers, two rules prevent most connection issues:
+
+1. A service that should be reachable from outside its own network namespace must listen on `0.0.0.0`.
+2. A container cannot use `localhost` to reach services running on your host machine.
+
+Examples:
+
+- If you run a server inside Docker and want to access it from your Mac browser, ensure the process binds to `0.0.0.0` inside the container (not `127.0.0.1`).
+- If a container needs to call a host service (for example Ollama running on macOS), use `host.docker.internal` from inside the container.
+
+Use this in containerized configs:
+
+```env
+OLLAMA_BASE_URL=http://host.docker.internal:11434
+```
+
+And ensure the host service is actually listening on an external interface (typically `0.0.0.0`) rather than only `127.0.0.1`.
+
+Quick reference:
+
+- Host -> Container (published port): `http://localhost:<published-port>`
+- Container -> Host service: `http://host.docker.internal:<port>`
+- Inside one container, `localhost` means that same container only
+
 ### 5. Create the .env file
 
 ```bash
@@ -131,15 +157,23 @@ WHISPER_CLI_PATH=../whisper.cpp/build/bin/whisper-cli
 WHISPER_MODEL_PATH=./models/kb_whisper_ggml_medium.bin
 WHISPER_SMALL_MODEL_PATH=./models/kb_whisper_ggml_small.bin
 
+# Shared ML backend switch (used by diarization and transcription): "torch" or "mlx"
+BACKEND=torch
+TORCH_DIARIZATION_MODEL=pyannote/speaker-diarization-3.1
+MLX_DIARIZATION_MODEL=mlx-community/diar_sortformer_4spk-v1-fp16
+
 STORAGE_PATH=./storage
 
-# Hugging Face token (needed for pyannote.audio speaker diarization)
+# Hugging Face token (needed only for the torch/pyannote diarization backend)
 # Get yours at https://huggingface.co/settings/tokens
 # You must accept the model terms at https://huggingface.co/pyannote/speaker-diarization-3.1
 HF_AUTH_TOKEN=hf_your_token_here
 EOF
 ```
 
+Set `BACKEND=mlx` to use MLX for diarization and transcription.
+Quick note: when using MLX, see the specific steps below for installing the python dependencies.
+
 Edit the file and fill in your actual paths and tokens.
 
 ### 6. Set up the Python backend
@@ -150,6 +184,12 @@ source venv/bin/activate
 pip install -r requirements.txt
 ```
 
+If you plan to run with `BACKEND=mlx`, install MLX-specific dependencies instead:
+
+```bash
+pip install -r requirements_mlx.txt
+```
+
 ### 7. Set up the frontend
 
 ```bash

api/actions.py

@@ -5,6 +5,7 @@
 from database import get_db
 from models import Action, ActionResult, ActionResultStatus, Meeting
 from tasks.action_task import run_action_task
+from translations import get_translation
 
 router = APIRouter(prefix="/api/actions", tags=["actions"])
 
@@ -40,7 +41,7 @@ def create_action(req: CreateActionRequest, db: Session = Depends(get_db)):
 def update_action(action_id: str, req: UpdateActionRequest, db: Session = Depends(get_db)):
     action = db.query(Action).filter(Action.id == action_id).first()
     if not action:
-        raise HTTPException(404, "Action not found")
+        raise HTTPException(404, get_translation('errors.action_not_found'))
     if req.name is not None:
         action.name = req.name
     if req.prompt is not None:
@@ -54,7 +55,7 @@ def update_action(action_id: str, req: UpdateActionRequest, db: Session = Depend
 def delete_action(action_id: str, db: Session = Depends(get_db)):
     action = db.query(Action).filter(Action.id == action_id).first()
     if not action:
-        raise HTTPException(404, "Action not found")
+        raise HTTPException(404, get_translation('errors.action_not_found'))
     db.delete(action)
     db.commit()
     return {"ok": True}
@@ -66,11 +67,11 @@ def delete_action(action_id: str, db: Session = Depends(get_db)):
 def run_action(action_id: str, meeting_id: str, db: Session = Depends(get_db)):
     action = db.query(Action).filter(Action.id == action_id).first()
     if not action:
-        raise HTTPException(404, "Action not found")
+        raise HTTPException(404, get_translation('errors.action_not_found'))
 
     meeting = db.query(Meeting).filter(Meeting.id == meeting_id).first()
     if not meeting:
-        raise HTTPException(404, "Meeting not found")
+        raise HTTPException(404, get_translation('errors.meeting_not_found'))
 
     result = ActionResult(
         action_id=action_id,
@@ -87,9 +88,10 @@ def run_action(action_id: str, meeting_id: str, db: Session = Depends(get_db)):
         db.commit()
     except Exception as e:
         result.status = ActionResultStatus.FAILED
-        result.error = f"Failed to queue task: {e}"
+        error_msg = get_translation('errors.failed_to_queue_task').format(error=e)
+        result.error = error_msg
         db.commit()
-        raise HTTPException(503, f"Task queue unavailable: {e}")
+        raise HTTPException(503, get_translation('errors.task_queue_unavailable').format(error=e))
 
     return result.to_dict()
 
@@ -112,7 +114,7 @@ def list_results(meeting_id: str, db: Session = Depends(get_db)):
     for r in results:
         d = r.to_dict()
         a = actions.get(r.action_id)
-        d["action_name"] = a.name if a else "Deleted action"
+        d["action_name"] = a.name if a else get_translation('ui.deleted_action')
         out.append(d)
     return out
 
@@ -121,7 +123,7 @@ def list_results(meeting_id: str, db: Session = Depends(get_db)):
 def delete_result(result_id: str, db: Session = Depends(get_db)):
     result = db.query(ActionResult).filter(ActionResult.id == result_id).first()
     if not result:
-        raise HTTPException(404, "Result not found")
+        raise HTTPException(404, get_translation('errors.result_not_found'))
     db.delete(result)
     db.commit()
     return {"ok": True}

api/analytics.py

@@ -3,6 +3,7 @@
 
 from database import get_db
 from models import Meeting, Speaker, Segment
+from translations import get_translation
 
 router = APIRouter(prefix="/api", tags=["analytics"])
 
@@ -11,7 +12,7 @@
 def get_meeting_analytics(meeting_id: str, db: Session = Depends(get_db)):
     meeting = db.query(Meeting).filter(Meeting.id == meeting_id).first()
     if not meeting:
-        raise HTTPException(404, "Meeting not found")
+        raise HTTPException(404, get_translation('errors.meeting_not_found'))
 
     speakers = db.query(Speaker).filter(Speaker.meeting_id == meeting_id).all()
     segments = (

api/encryption.py

@@ -9,6 +9,7 @@
 from models.segment import Segment
 from models.action import ActionResult
 from services.encryption_service import EncryptionService
+from translations import get_translation
 
 router = APIRouter(prefix="/api/meetings", tags=["encryption"])
 
@@ -26,11 +27,11 @@ class DecryptRequest(BaseModel):
 def encrypt_meeting(meeting_id: str, req: EncryptRequest, db: Session = Depends(get_db)):
     meeting = db.query(Meeting).filter(Meeting.id == meeting_id).first()
     if not meeting:
-        raise HTTPException(404, "Meeting not found")
+        raise HTTPException(404, get_translation('errors.meeting_not_found'))
     if meeting.is_encrypted:
-        raise HTTPException(400, "Meeting is already encrypted")
+        raise HTTPException(400, get_translation('errors.meeting_already_encrypted'))
     if not req.password:
-        raise HTTPException(400, "Password is required")
+        raise HTTPException(400, get_translation('errors.password_required'))
 
     svc = EncryptionService
     salt = svc.generate_salt()
@@ -66,13 +67,13 @@ def encrypt_meeting(meeting_id: str, req: EncryptRequest, db: Session = Depends(
 def decrypt_meeting(meeting_id: str, req: DecryptRequest, db: Session = Depends(get_db)):
     meeting = db.query(Meeting).filter(Meeting.id == meeting_id).first()
     if not meeting:
-        raise HTTPException(404, "Meeting not found")
+        raise HTTPException(404, get_translation('errors.meeting_not_found'))
     if not meeting.is_encrypted:
-        raise HTTPException(400, "Meeting is not encrypted")
+        raise HTTPException(400, get_translation('errors.meeting_not_encrypted'))
 
     svc = EncryptionService
     if not svc.check_password(req.password, meeting.encryption_salt, meeting.encryption_verify):
-        raise HTTPException(403, "Wrong password")
+        raise HTTPException(403, get_translation('errors.wrong_password'))
 
     salt = base64.b64decode(meeting.encryption_salt)
     key = svc.derive_key(req.password, salt)

api/export.py

@@ -8,11 +8,10 @@
 from database import get_db
 from models import Meeting, Segment
 from models.action import Action, ActionResult
+from translations import get_translation
 
 router = APIRouter(prefix="/api", tags=["export"])
 
-UNKNOWN_SPEAKER = "Okand"
-
 
 def _safe_filename(name: str) -> str:
     """Sanitize a string for use in Content-Disposition filename."""
@@ -54,7 +53,7 @@ def export_meeting(
 ):
     meeting = db.query(Meeting).filter(Meeting.id == meeting_id).first()
     if not meeting:
-        raise HTTPException(404, "Meeting not found")
+        raise HTTPException(404, get_translation('errors.meeting_not_found'))
 
     segments = (
         db.query(Segment)
@@ -79,13 +78,13 @@ def export_meeting(
     elif format == "pdf":
         return _export_pdf(meeting, segments)
     else:
-        raise HTTPException(400, f"Unknown format: {format}")
+        raise HTTPException(400, get_translation('errors.unknown_format').format(format=format))
 
 
 def _export_srt(meeting: Meeting, segments: list[Segment]) -> PlainTextResponse:
     lines = []
     for i, seg in enumerate(segments, 1):
-        speaker = seg.speaker.display_name if seg.speaker else UNKNOWN_SPEAKER
+        speaker = seg.speaker.display_name if seg.speaker else get_translation('common.unknown_speaker')
         lines.append(str(i))
         lines.append(f"{format_srt_time(seg.start_time)} --> {format_srt_time(seg.end_time)}")
         lines.append(f"[{speaker}] {seg.text}")
@@ -103,7 +102,7 @@ def _export_srt(meeting: Meeting, segments: list[Segment]) -> PlainTextResponse:
 def _export_vtt(meeting: Meeting, segments: list[Segment]) -> PlainTextResponse:
     lines = ["WEBVTT", ""]
     for seg in segments:
-        speaker = seg.speaker.display_name if seg.speaker else UNKNOWN_SPEAKER
+        speaker = seg.speaker.display_name if seg.speaker else get_translation('common.unknown_speaker')
         lines.append(f"{format_vtt_time(seg.start_time)} --> {format_vtt_time(seg.end_time)}")
         lines.append(f"<v {speaker}>{seg.text}")
         lines.append("")
@@ -121,7 +120,7 @@ def _export_txt(meeting: Meeting, segments: list[Segment]) -> PlainTextResponse:
     lines = []
     current_speaker = None
     for seg in segments:
-        speaker = seg.speaker.display_name if seg.speaker else UNKNOWN_SPEAKER
+        speaker = seg.speaker.display_name if seg.speaker else get_translation('common.unknown_speaker')
         if speaker != current_speaker:
             if lines:
                 lines.append("")
@@ -148,7 +147,7 @@ def _export_json(meeting: Meeting, segments: list[Segment]) -> JSONResponse:
             {
                 "start": seg.start_time,
                 "end": seg.end_time,
-                "speaker": seg.speaker.display_name if seg.speaker else UNKNOWN_SPEAKER,
+                "speaker": seg.speaker.display_name if seg.speaker else get_translation('common.unknown_speaker'),
                 "text": seg.text,
             }
             for seg in segments
@@ -170,7 +169,7 @@ def _export_md(meeting: Meeting, segments: list[Segment]) -> PlainTextResponse:
 
     current_speaker = None
     for seg in segments:
-        speaker = seg.speaker.display_name if seg.speaker else UNKNOWN_SPEAKER
+        speaker = seg.speaker.display_name if seg.speaker else get_translation('common.unknown_speaker')
         if speaker != current_speaker:
             lines.append("")
             ts = format_timestamp_short(seg.start_time)
@@ -204,7 +203,7 @@ def _export_docx(meeting: Meeting, segments: list[Segment]) -> StreamingResponse
 
     current_speaker = None
     for seg in segments:
-        speaker = seg.speaker.display_name if seg.speaker else UNKNOWN_SPEAKER
+        speaker = seg.speaker.display_name if seg.speaker else get_translation('common.unknown_speaker')
         if speaker != current_speaker:
             ts = format_timestamp_short(seg.start_time)
             doc.add_heading(f"{speaker} [{ts}]", level=3)
@@ -272,7 +271,7 @@ def _export_pdf(meeting: Meeting, segments: list[Segment]) -> StreamingResponse:
 
     current_speaker = None
     for seg in segments:
-        speaker = seg.speaker.display_name if seg.speaker else UNKNOWN_SPEAKER
+        speaker = seg.speaker.display_name if seg.speaker else get_translation('common.unknown_speaker')
         if speaker != current_speaker:
             ts = format_timestamp_short(seg.start_time)
             story.append(Paragraph(f"{speaker} [{ts}]", speaker_style))
@@ -306,15 +305,15 @@ def export_action_result(
 ):
     result = db.query(ActionResult).filter(ActionResult.id == result_id).first()
     if not result:
-        raise HTTPException(404, "Action result not found")
+        raise HTTPException(404, get_translation('errors.action_result_not_found'))
     if not result.result_text:
-        raise HTTPException(400, "Action result has no content")
+        raise HTTPException(400, get_translation('errors.action_result_has_no_content'))
 
     action = db.query(Action).filter(Action.id == result.action_id).first()
-    action_name = action.name if action else "Action"
+    action_name = action.name if action else get_translation('ui.action')
 
     meeting = db.query(Meeting).filter(Meeting.id == result.meeting_id).first()
-    meeting_title = meeting.title if meeting else "Meeting"
+    meeting_title = meeting.title if meeting else get_translation('ui.meeting')
 
     filename = _safe_filename(f"{meeting_title} - {action_name}")
 
@@ -324,7 +323,7 @@ def export_action_result(
             headers={"Content-Disposition": f'attachment; filename="{filename}.txt"'},
         )
     elif format == "md":
-        md_content = f"# {action_name}\n\n**Meeting:** {meeting_title}\n\n---\n\n{result.result_text}"
+        md_content = f"# {action_name}\n\n**{get_translation('ui.meeting')}:** {meeting_title}\n\n---\n\n{result.result_text}"
         return PlainTextResponse(
             md_content,
             headers={
@@ -337,7 +336,7 @@ def export_action_result(
     elif format == "pdf":
         return _export_action_pdf(result.result_text, action_name, meeting_title, filename)
     else:
-        raise HTTPException(400, f"Unknown format: {format}")
+        raise HTTPException(400, get_translation('errors.unknown_format').format(format=format))
 
 
 def _export_action_docx(text: str, action_name: str, meeting_title: str, filename: str) -> StreamingResponse:
@@ -347,7 +346,7 @@ def _export_action_docx(text: str, action_name: str, meeting_title: str, filenam
     doc = Document()
     doc.add_heading(action_name, level=1)
     p = doc.add_paragraph()
-    run = p.add_run(f"Meeting: {meeting_title}")
+    run = p.add_run(f"{get_translation('ui.meeting')}: {meeting_title}")
     run.font.size = Pt(10)
     run.font.color.rgb = RGBColor(128, 128, 128)
 
@@ -391,7 +390,7 @@ def _export_action_pdf(text: str, action_name: str, meeting_title: str, filename
 
     story = []
     story.append(Paragraph(action_name, title_style))
-    story.append(Paragraph(f"Meeting: {meeting_title}", meta_style))
+    story.append(Paragraph(f"{get_translation('ui.meeting')}: {meeting_title}", meta_style))
 
     for line in text.split("\n"):
         safe = line.replace("&", "&amp;").replace("<", "&lt;").replace(">", "&gt;")