Simple proof-of-concept Flask chat application created to test the @groq inference API free-tier using the surprisingly capable qwen/qwen3-32b model.
Besides AI-powered chat, speech capabilities are integrated, in particular using whisper-large-v3-turbo (also hosted @groq) for speech-to-text (STT) and the package edge-tts for text-to-speech (TTS).
Dockerized for reproducibility, ease of use and deployment. Keeps seperate sessions for separate users with recoverable chat logs in-between deployments. Uses gunicorn to serve the Flask application and valkey (an open community Redis fork) to store session histories in append-only files.
相棒は色々な使い方があります。 例えば、日本語を練習するためにとても良いパトナです。 ぜひ、試してみてください。
Aibou serves on localhost:4200 by default.
Currently STT and TTS are offloaded to external providers, but in theory, small, locally runnable models such as Qwen3-TTS-0.6B model could be used to replace edge-tts.
Perhaps there are even equivalent edge STT models that could be utilized.
TTS must currently be configured to the correct target language beforehand config.py.
Since the TTS voice parameter is passed every request, simple language detection could be used to select an appropriate voice output.
For consistent cross-language voice output, a voice-cloned omni-lingual TTS model becomes necessary.
STT is currently also hinted towards Japanese, but STT works fine even if other languages are spoken.
Separate sentiment analysis requests on user-generated prompts could be used and bundled within requests to keep output tone consistent. In the future, this might even be used to instruct the TTS model on what tone to speak in.
Currently STT result text is rendered only after the user finished speaking. In a future iteration, text could be streamed as it is being detected. On a related note, no analysis is performed to detect end-of-speech, instead this has to be triggered manually.
Permissions for persistent data created in valkey-data might need to be reconfigured if the container is rebuilt, for instance by running docker-compose down && docker-compose up --build.
But if the container is not rebuilt, data can be persisted just fine