Whisper Transcription Server

Local speech-to-text via whisper.cpp with GPU acceleration on port 5555. Accepts WAV/MP3 and other common formats (16kHz mono WAV is optimal) as multipart POST.

Use it

curl -X POST http://localhost:5555/inference -F "[email protected]"

Returns JSON {"text": ...}.

Server

• Binary: ~/whisper.cpp/build/bin/whisper-server
• Model: ~/whisper.cpp/models/ggml-large-v3.bin
• Port: 5555
• Runs as the whisper-server systemd service (systemctl status whisper-server, logs via journalctl -u whisper-server -f).

Systemd unit at /etc/systemd/system/whisper-server.service:

[Unit]
Description=Whisper.cpp Transcription Server
After=network.target

[Service]
Type=simple
User=matt
WorkingDirectory=/home/matt/whisper.cpp/build
ExecStart=/home/matt/whisper.cpp/build/bin/whisper-server \
  -m /home/matt/whisper.cpp/models/ggml-large-v3.bin \
  -l en \
  --port 5555 \
  --host 0.0.0.0 \
  --threads 4
Restart=always
RestartSec=10

[Install]
WantedBy=multi-user.target

Performance / VRAM

~2–4x real-time; the first request after startup is slow (model load). large-v3 uses ≈ 6GB VRAM on the shared RTX 3090 — on OOM, wait for other GPU services to unload (see the Vram-GPU-OOM skill for the retry/signaling pattern).