Hi, I'm Sakharam Thorat 👨‍💻
AI Voice Systems Engineer | Real-time Speech AI | Conversational Agents

🚀 I specialize in building production-grade real-time voice AI systems — from telephony pipelines to speech-to-speech agents.

My work spans across:

• 🎙️ Speech AI (STT, TTS, Speech-to-Speech)
• ⚡ Low-latency real-time systems
• 📡 SIP / VoIP / streaming architectures
• 🧠 Voice agents & conversational AI
• 🧩 On-device AI & optimization

• 📍 Based in Pune, India
• 💼 10+ years in real-time communication systems (Vonage, Avaya)
• 🔊 Built end-to-end voice AI pipelines (STT → LLM → TTS)
• ⚡ Focused on low-latency conversational systems (<200ms–2s)
• 🧠 Exploring AI voice agents, speech models & edge AI

📧 Reach me at: srt.2011@outlook.com
🔗 LinkedIn

• Real-time speech-to-speech AI (Azure OpenAI)
• Multi-provider STT (Deepgram, Google, Azure, Whisper)
• TTS pipelines (Edge, streaming synthesis)
• Voice agents with multi-turn conversations

• SIP / RTP / VoIP pipelines
• WebRTC + WebSocket streaming
• Codec optimization (G.711, G.729, Opus)
• Packetization (ptime), jitter & packet loss handling

• Noise suppression (RNN-based training on A100)
• CTC-based ASR experimentation
• NVIDIA NeMo, Vosk (Kaldi-based)
• Audio quality evaluation (PESQ, MOS)

• Whisper.cpp, Parakeet (on-device STT)
• Model quantization for low-memory inference
• CPU-optimized real-time pipelines

• aura-sde-interview-agent
  Real-time voice AI agent for Google SDE interviews
  → Full loop: Behavioural, Coding, System Design, Debugging
  → Built with Gemini Live + Vertex AI
  → Live grading + voice-only interaction

• nemotron-stt (In Progress)
  High-concurrency WebSocket ASR server using Nemotron-0.6B

• whisper-stt
  Production-grade streaming STT using faster-whisper turbo

• qwen3-stt
  Voice AI agent design + STT experimentation

• VeloxTx (In Progress)
  Multilingual low-latency translation engine
  → “Train anywhere, run anywhere” philosophy

• velox-realtime (In Progress)
  CPU-optimized real-time ASR (<30ms latency)

• x-benchmark-tests
  Benchmarking STT models (accuracy, latency, throughput)

• stt-systems-lab (In Progress)
  Experimental lab for:

  • Nemotron
  • Whisper Turbo
  • Qwen3
  • ONNX / browser inference

• freeswitch-speech-ai
  Real-time:

  • Transcription
  • Translation
  • Sentiment
  • NER
  • Call summarization

• pjsua2-python
  Prebuilt Python bindings for PJSIP (SIP stack)

• slm_framework_full (In Progress)
  Unified framework for:
  • STT, TTS, Vocoder
  • Small Language Models
  • Real-time inference

• STT: Whisper, Deepgram, Azure, Google
• TTS: Edge, streaming pipelines
• Speech Models: NeMo, Vosk, Kaldi-based
• Metrics: WER, PESQ, MOS

• FastAPI, WebSockets, WebRTC
• LiveKit, Cloud Run, Vertex AI
• Kubernetes, Docker

• Python (primary)
• TypeScript / JavaScript
• Java

• Real-time AI voice agents with tool-calling
• Ultra-low latency speech-to-speech systems
• On-device conversational AI (edge-first)
• Production-grade multilingual voice pipelines

• 💼 LinkedIn
• 📧 srt.2011@outlook.com

⭐ If you're working on Voice AI / Speech Systems / Real-time AI — let’s collaborate!