Project ideas from Hacker News discussions.

Launch HN: Speko (YC S26) – OpenRouter for Voice AI

📝 Discussion Summary (Click to expand)

3 Most Prevalent Themes

Theme Summary Representative Quote
1. Automated benchmarking & routing are the core value proposition Users stress that the real benefit of the platform is its continuous measurement of voice‑model performance and smart routing based on those results, rather than just offering a set of models. The main difference from gateway is we help with picking the right voice stack… we benchmark the models continuously and route based on those measurements for your language and constraints…” – abdik
2. Production voice agents are still mostly cascaded, not end‑to‑end Enterprise‑grade deployments continue to rely on separate STT, LLM, and TTS components; full‑stack end‑to‑end models have not yet replaced this architecture at scale. In my experience, this isn't true… enterprises are not moving towards STS solutions in a meaningful way…” – cootsnuck
exactly, we see the same thing, around 95% cases are still cascaded…” – abdik
3. Turn‑taking and integrated “conversation‑in‑a‑box” APIs are essential, with growing interest in on‑device/local models There is strong demand for a single API that handles VAD, turn‑taking, and session state, and many commenters discuss the future of on‑device inference versus hosted services. Yes, on the hosted side (agents platform): full sessions come with VAD and turn‑taking handled – we set them up and tune them for your use case, so that is the closest thing to conversation in a box.” – abdik

🚀 Project Ideas

VoiceMesh Router

Summary

  • A self‑hosted routing layer that automatically selects the best voice model (STT, TTS, LLM) for a call based on real‑time benchmark scores.
  • Eliminates vendor lock‑in and reduces latency while providing built‑in turn‑taking orchestration.

Details

Key Value
Target Audience Developers building scalable voice agents who need multi‑provider flexibility
Core Feature Dynamic model routing with integrated turn‑taking API
Tech Stack Docker + FastAPI + Rust inference; Redis caching; OpenTelemetry metrics
Difficulty Medium
Monetization Revenue-ready: SaaS‑hosted control plane ($0.01 per routed turn)

Notes

  • HN commenters asked for a "conversation in a box" and public benchmarks – this delivers both.
  • Addresses the need for a single API that handles turn‑taking and swaps models automatically.

VoiceArena Dashboard

Summary

  • A hosted evaluation platform that scores voice models across STT, TTS, and LLM using automated latency, WER/CER, and style‑similarity metrics.
  • Provides an “LM Arena for voice” so teams can compare and switch models without manual testing.

Details

Key Value
Target Audience Product managers and AI teams evaluating voice components
Core Feature Unified benchmarks + automatic style‑prompt matching for seamless model swapping
Tech Stack Python backend, React frontend, PostgreSQL, HuggingFace models
Difficulty High
Monetization Revenue-ready: Tiered subscription (Starter $49/mo, Pro $299/mo)

Notes

  • Directly answers abdik’s point that “automatic evals” are the real value.
  • Users want to switch providers while keeping output style consistent – VoiceArena makes that painless.

DomainSTT Studio

Summary

  • No‑code portal to fine‑tune and prompt‑augment STT models with custom domain vocabularies, reducing transcription errors for niche terms.
  • Automates continuous WER monitoring to keep domain‑specific accuracy high.

Details

Key Value
Target Audience Enterprises in regulated domains (legal, medical, finance)
Core Feature Domain‑specific keyword prompt injection + auto‑re‑training pipeline
Tech Stack Streamlit UI, HuggingFace Whisper fine‑tuning, AWS SageMaker, GitHub Actions CI/CD
Difficulty Low
Monetization Hobby

Notes

  • Mirrors user complaints about “Claude Code” being transcribed as “Cloud Code” and the need for prompt‑based vocabulary.
  • Aligns with discussions on synthetic data fine‑tuning and domain‑specific STT improvements.

Read Later