Original version of WhisperLive compose.yaml & whisper-live.env

2026-07-07 15:51:04 -04:00
parent b75e6bb2ab
commit 543af6f9ba
2 changed files with 70 additions and 0 deletions
@@ -0,0 +1,25 @@
+services:
+  whisper-live:
+    image: hwdsl2/whisper-live-server:latest
+    container_name: whisper-live
+    restart: unless-stopped
+    ports:
+      - "9097:9090"   # WebSocket — real-time streaming clients
+      - "8071:8000"   # REST API — OpenAI-compatible /v1/audio/transcriptions
+    volumes:
+      - whisper-live-data:/var/lib/whisper-live   # persists downloaded models
+    env_file:
+      - whisper-live.env
+    # --- ADDED FOR INTEL iGPU PASSTHROUGH ---
+    devices:
+      - /dev/dri/renderD128:/dev/dri/renderD128
+      - /dev/dri/card1:/dev/dri/card1
+    # ----------------------------------------
+
+    # The 13700K has 8 P-cores; give the container access to all of them.
+    # Adjust cpuset if you want to reserve cores for other workloads.
+    cpuset: "0-15"          # all 16 logical P-core threads
+    mem_limit: 6g           # medium int8 needs ~2 GB; headroom for concurrency
+
+volumes:
+  whisper-live-data:
@@ -0,0 +1,45 @@
+# whisper-live.env
+# Configuration for hwdsl2/whisper-live-server on Intel Core i9-13700K (CPU-only)
+# Copy this file to the same directory as docker-compose.yml before starting.
+
+# --- Model ---
+# 'medium.en' is the sweet spot for CPU-only English meetings:
+# accurate, comfortably real-time, ~1.5 GB RAM.
+# Upgrade to 'large-v3-turbo' if accuracy is more important than latency;
+# downgrade to 'small.en' if you find CPU usage too high.
+WHISPERLIVE_MODEL=medium.en
+
+# --- Language ---
+# Hardcode to 'en' to skip autodetection — saves ~100 ms per chunk.
+# Change to 'auto' if your meetings include other languages.
+WHISPERLIVE_LANGUAGE=en
+
+# --- Ports ---
+WHISPERLIVE_PORT=9090        # WebSocket (streaming clients)
+WHISPERLIVE_REST_PORT=8000   # REST API
+
+# --- Concurrency ---
+# 1 simultaneous session is safest for real-time CPU-only use.
+# Increase to 2 cautiously — each session competes for the same cores.
+WHISPERLIVE_MAX_CLIENTS=1
+
+# --- Session length ---
+# 7200 seconds = 2 hours, suitable for a full meeting recording.
+WHISPERLIVE_MAX_CONNECTION_TIME=7200
+
+# --- Voice Activity Detection ---
+# Keep enabled: VAD skips silence, which meaningfully reduces CPU load
+# during pauses and cross-talk at a 12-person table.
+WHISPERLIVE_USE_VAD=true
+
+# --- CPU threads ---
+# The 13700K has 8 P-cores. Setting threads to 8 gives best single-session
+# latency. If you raise MAX_CLIENTS to 2, drop this to 4.
+WHISPERLIVE_THREADS=8
+
+# --- Logging ---
+WHISPERLIVE_LOG_LEVEL=INFO
+
+# --- Backend ---
+WHISPER_BACKEND=openvino
+