dubbing-tts

The TTS weights the Blaze dubbing service synthesises with. Currently the Khmer OmniVoice fine-tune, checkpoint-4000 of omnivoice_finetune_khmer. This was the only copy that existed outside one machine's disk.

The repo is named for the role, not the model, so swapping the engine or adding a language does not mean re-pointing every build: services/tts/Dockerfile pulls whatever is on the branch it is told to.

Served over HTTP by services/tts (OmniVoice sidecar), never loaded in-process. Measured throughput on an L4: 23.6 lines/min on one replica, 28.9 on two replicas on separate cards. Two replicas sharing a card is worse than one (15.7/min) and then OOMs mid-generation, because a replica needs ~1.2 GB of transient headroom.

Branches

main is what image builds pin. dev is where a new fine-tune lands so it can be A/B'd against main before promotion:

HF_REVISION=dev docker compose --profile tts build tts