English-German GPU speech translation with cloud LLM, memory and voice interruption.
279
GitHub source, demo and full documentation
Talk in English, German or both. The application supports conversation, confirmed translation-mode changes, teaching, persistent notes and voice interruption. Qwen3-ASR-0.6B recognizes speech locally; Silero VAD handles turn timing; Ollama Cloud Gemma 4 31B generates streamed replies; Kokoro English and Piper Thorsten German produce ordered audio chunks locally.
Requires Linux x86-64 containers and an NVIDIA GPU exposed to Docker. The tested native GPU has 6 GiB VRAM. Verified GPU/container release, digest and measurements. This is not a CPU image.
Create a local .env containing your own credentials:
OLLAMA_API_KEY=your_own_key
OLLAMA_BASE_URL=https://ollama.com
OLLAMA_MODEL=gemma4:31b
docker pull bmaged23/linguabridge-en-de-sts:gpu
docker run --rm --gpus all --env-file .env \
-p 127.0.0.1:8766:8766 \
-v linguabridge-memory:/app/data \
bmaged23/linguabridge-en-de-sts:gpu
After warmup, open http://localhost:8766 and allow microphone access. No Git clone or first-run model download is needed: the image contains the production speech models. The cloud LLM requires your own account allowance. Your microphone stays in the browser, so no audio device passthrough is required. Notes/history persist in the named Docker volume. Stop with Ctrl+C.
The FastAPI backend exposes HTTP health/state/configuration/control endpoints and WS /ws for streaming microphone audio, STT text, LLM replies and synthesized speech. Start the container first, then connect at http://localhost:8766. See the API README for endpoint examples, audio formats, playback acknowledgements and the current same-origin, single-session scope.
The qualified image passed 94 offline tests, 52/52 STT recordings and 17/17 browser turns, with zero JavaScript errors. Median first audio after recording ended was 2.53 seconds. Highest observed container memory was 2,601 MiB and whole-device VRAM 4,954 MiB on a GTX 1660 Ti with 6 GiB VRAM; the cloud LLM hardware is excluded. These are sampled observations on one setup, not device-wide guarantees. Whisper independently scored 37/37 output chunks; English agreement WER/CER 4.04%/2.84%, German 9.09%/7.08%, including number-format differences and one lexical disagreement. See the linked report for all comparisons and limitations.
Immutable image: bmaged23/linguabridge-en-de-sts@sha256:04ec957cf2fb6df7e0d9b0ca378ac7f3f37e5fcc79ebcd04181ce36b2c08bcbd.
The candidate image runs offline tests, warms real GPU models, compares 52 archived ASR waveforms against frozen references with per-recording WER/CER and meaning/routing gates, and completes 17 prerecorded browser conversation turns. Publication follows passing gates. The workflow pulls the published tag, verifies its identity and tests normal startup again. GitHub-hosted tests and the trusted GPU evaluation runner have separate roles.
Native measurements: caching completed ASR segments with GPU batches of two preserved all 52 final transcripts. A matched mixed story's first-audio delay fell from 7.77 to 2.69 seconds; the 17-turn native suite had 2.62-second median delay. These are native measurements, not a container performance guarantee. Independent recognition scored 37 generated chunks without execution failures; number/time formatting explains the reviewed text/audio differences. Noise and human code-switching remain limitations.
See the evaluation reports and model notices. The image excludes personal memory, recordings, raw test inputs and API keys. Finalized text and relevant memory go to Ollama Cloud; recognition and speech synthesis run locally.
Content type
Image
Digest
sha256:04ec957cf…
Size
5.4 GB
Last updated
25 days ago
docker pull bmaged23/linguabridge-en-de-sts:gpu