Real-Time Waiter Voice Identification
Budget: $30 – $250 USD
I need a standalone Python module (FastAPI preferred) that can single out which waiter is speaking in a busy restaurant and return the correct Waiter_ID in well under 300 ms.
Live audio will arrive directly from a microphone stream, so the system has to stay robust against kitchen clatter, background music and overlapping chatter.
Core flow I have in mind
• Extract speaker embeddings with a model such as ECAPA-TDNN or Pyannote and store them efficiently (SQLite is fine, a vector DB is even better).
• Expose an enrollment endpoint that my mobile app can hit automatically whenever a new waiter records their sample; the call should save the embedding to the tenant-specific store.
• Expose a real-time identify endpoint that accepts short chunks, compares them to the active voice set and responds with Waiter_ID plus a confidence score.
Multi-tenant readiness is important: each restaurant uses its own isolated voice database, yet the codebase remains a single deployment.
Acceptance criteria
– Median end-to-end latency ≤ 300 ms on a mid-range CPU.
– Identification accuracy acceptable for production dining rooms with moderate noise (we can validate on a sample set together).
– Clean Python code, requirements.txt, and concise README describing setup, model weights, and how to extend the storage layer.
If you have previous SID or audio-ML benchmarks to share, that will help me choose quickly.
Live audio will arrive directly from a microphone stream, so the system has to stay robust against kitchen clatter, background music and overlapping chatter.
Core flow I have in mind
• Extract speaker embeddings with a model such as ECAPA-TDNN or Pyannote and store them efficiently (SQLite is fine, a vector DB is even better).
• Expose an enrollment endpoint that my mobile app can hit automatically whenever a new waiter records their sample; the call should save the embedding to the tenant-specific store.
• Expose a real-time identify endpoint that accepts short chunks, compares them to the active voice set and responds with Waiter_ID plus a confidence score.
Multi-tenant readiness is important: each restaurant uses its own isolated voice database, yet the codebase remains a single deployment.
Acceptance criteria
– Median end-to-end latency ≤ 300 ms on a mid-range CPU.
– Identification accuracy acceptable for production dining rooms with moderate noise (we can validate on a sample set together).
– Clean Python code, requirements.txt, and concise README describing setup, model weights, and how to extend the storage layer.
If you have previous SID or audio-ML benchmarks to share, that will help me choose quickly.