Prepare Local Speech Recognition for an Offline Event

Start the forallcli sidecar, choose Qwen3 ASR, enumerate audio, and preflight local recognition.

Written By 4ALL.LIVE

Last updated About 1 month ago

Prepare local speech recognition so audio can be processed on the operator workstation without the cloud STT runtime once required local resources are available.

Best for: Desktop operators and technical administrators.

Before you start

Requires the desktop app, compatible hardware, local/offline entitlement where applicable, and downloaded local resources.

  • Use a non-production event for commissioning.
  • Record the current build and workstation owner.
  • Prepare a documented fallback before making changes.
  • Use a production-capable workstation with adequate CPU/GPU, memory, storage, cooling, and power.
  • Download all required resources before the venue loses internet.
  • Prepare the exact microphone, channel, language, and custom vocabulary.

Planning guide

  1. Open Event Setup and set Speech engine to Offline (legacy label Whisper Local/Edge maps to Offline).
  2. The desktop app initializes the bundled forallcli sidecar on loopback port 8787.
  3. Confirm /health succeeds and audio devices load from /api/audio/devices.
  4. Select Qwen3-ASR-1.7B-bf16 (repository model ID mlx-community/Qwen3-ASR-1.7B-bf16).
  5. Select the stable device UID and zero-based input channel; default sample rate for the local request is 16 kHz.
  6. Choose language, decide whether on-device diarization is enabled, and load custom vocabulary if required.
  7. Start a rehearsal and verify connecting → connected → running, partial transcripts, final transcripts, speaker attribution, and stop.
  8. Disconnect the external network in a controlled test and verify the exact offline scope before travel.

What success looks like: The local sidecar recognizes the intended microphone and produces partial/final captions without relying on cloud STT during the tested scenario.

Check your setup

  • Sidecar health and WebSocket ws://127.0.0.1:8787/ws/realtime work.
  • The model is available locally.
  • Correct device/channel are used.
  • Thermals and latency stay acceptable through a sustained rehearsal.

Troubleshooting

  • No devices: restart sidecar/app and regrant microphone access.
  • Health fails: endpoint security may block forallcli or port 8787.
  • Model load fails: reconnect and complete/cache the resource before offline operation.
  • High latency: reduce competing load and use supported hardware.

Security and operational notes

Important: Offline STT does not automatically make the entire SaaS workflow offline. Sharing, translation, account checks, remote control, and synchronization may still require network services.