Speakr v0.10.8-alpha

This release fixes an error logged at startup with whisperx-asr-service (#409), makes every transcription request name the configured model, and fixes the creation of the admin account from the example configuration.

Fixed

• "Invalid model size ''" at startup with whisperx-asr-service. At startup, a short bundled clip is transcribed to identify the voice embedding model. No model name was sent with that request, so the service's own default model was used, and that default is empty when PRELOAD_MODEL is not set in the service's .env. The same model as an ordinary upload is now sent: the default transcription model from the admin settings, when one is set. If the check cannot run, the log now states that the message comes from this check, that no recording was affected, and, for a missing model, which setting to add. Reported by @cpuks (#409).
• Admin account from the example configuration. Every example configuration sets ADMIN_EMAIL=admin@example.com. That domain does not receive mail, so the address was rejected and a container started from an unchanged example stopped. The admin address is now checked for its format only; self-registration keeps the full check.
• Model settings in incognito mode and bulk reprocessing. In incognito mode, the default transcription model is used as for an upload. When the transcription of several recordings is reprocessed at once, each recording's tag and folder settings (model, hotwords and speaker counts) now apply, as for a single recording.
• Empty model names are no longer sent to an ASR service.
• API v1 GET /transcription returned an error when no model list had been saved in the admin settings. It now also reports the default model that transcriptions actually use.

Tests for the documented setup

The #409 error was missed because every test ran with a fully configured environment. New tests start Speakr with only the settings from config/env.whisperx.example and check every transcription request against the model rules of whisperx-asr-service, including the case in which the service has no default model. A new CI job runs Speakr against the real whisperx-asr-service code, on stand-in ML packages and without a GPU, from startup through an upload and a joined upload, and fails on any error in either log.

Upgrading

Pull the new image and start as usual. No database or configuration changes are required. If your whisperx-asr-service has no PRELOAD_MODEL, either set it (for example PRELOAD_MODEL=large-v3) or choose a default transcription model in Speakr's admin settings. With whisperx-asr-service 0.4.2, large-v3 is used when the variable is empty.

Full Changelog: v0.10.7-alpha...v0.10.8-alpha