STT, TTS , LLM Video Autoreply System
Budget: $30 – $250 USD
I need a lightweight, end-to-end solution that listens to spoken Indonesian during a live video call, converts the speech to text, passes that text through a small language model to craft an automatic reply, and then speaks the response back in a natural-sounding English voice.
Scope and key points
• Speech-to-Text: real-time transcription of incoming audio (Whisper, Google Speech, or a comparable engine).
• Language Model: compact LLM (OpenAI API, Llama.cpp, or similar) that turns the transcript into a concise reply.
• Text-to-Speech: natural voice rendering—no robotic tone—ready to play back with minimal latency.
• Pipeline: audio in ➜ STT ➜ LLM ➜ TTS ➜ audio out, all wrapped in a script or small service that I can trigger during a call.
• Compatibility: desktop environment (Windows or Linux); bonus if it can hook into common video-meeting tools via virtual audio cable or WebRTC.
• Deliverables: source code, quick-start guide, and a short demo video proving the loop works in real time.
Timing is critical—I’m aiming to see a working first cut as soon as possible, then iterate quickly if tweaks are needed. If you’ve built anything similar, I’d love to hear about it when you reply.
Scope and key points
• Speech-to-Text: real-time transcription of incoming audio (Whisper, Google Speech, or a comparable engine).
• Language Model: compact LLM (OpenAI API, Llama.cpp, or similar) that turns the transcript into a concise reply.
• Text-to-Speech: natural voice rendering—no robotic tone—ready to play back with minimal latency.
• Pipeline: audio in ➜ STT ➜ LLM ➜ TTS ➜ audio out, all wrapped in a script or small service that I can trigger during a call.
• Compatibility: desktop environment (Windows or Linux); bonus if it can hook into common video-meeting tools via virtual audio cable or WebRTC.
• Deliverables: source code, quick-start guide, and a short demo video proving the loop works in real time.
Timing is critical—I’m aiming to see a working first cut as soon as possible, then iterate quickly if tweaks are needed. If you’ve built anything similar, I’d love to hear about it when you reply.