Speech & Transcription

Hosted STT (speech-to-text)

• ★ Deepgram — fast, accurate, generous free credit ($200) on signup. Realtime + batch APIs. The default for production STT in 2026.
• ★ AssemblyAI — very accurate, great speaker diarization; free tier exists.
• OpenAI Whisper API — easy if you're already on OpenAI; cheaper batch than realtime.
• Google Speech-to-Text — long-time baseline, paid.
• AWS Transcribe — bundled with AWS.
• Azure Speech — bundled with Azure.
• Soniox, Speechmatics, Rev AI, Gladia — strong alternatives with various free tiers.

Self-host / local STT

• ★ whisper.cpp — fastest CPU/GPU port of Whisper; runs on a M1 Mac or a Hetzner box.
• WhisperX — Whisper + word-level timestamps + diarization.
• Vosk — offline, lightweight, real-time-friendly; many languages.
• Coqui STT (formerly Mozilla DeepSpeech) — open source.
• NeMo / Riva (NVIDIA) — heavyweight, GPU-required.
• Faster Whisper — CTranslate2-based; meaningfully faster than vanilla Whisper.

In-browser STT

• ★ Web Speech API — built into browsers; free; quality varies; great for prototypes.
• Transformers.js — Whisper in the browser via WebGPU; free, no server.
• WebLLM-style WASM — for offline-capable PWAs.

Realtime / streaming

• Deepgram Streaming, AssemblyAI Realtime, Speechmatics Realtime — websocket APIs.
• Soniox Realtime — competitive realtime accuracy.
• OpenAI Realtime API — speech-in / speech-out for voice agents.

Voice activity detection (VAD)

• Silero VAD — small ONNX model; great for "is the user speaking?".
• @ricky0123/vad-web — browser VAD wrapper.
• @ricky0123/vad-node — Node VAD.

Text-to-speech (TTS)

• ★ ElevenLabs — best voices; generous free tier; paid for production.
• OpenAI TTS — cheap, decent voices, easy if already on OpenAI.
• Cartesia — fast, low-latency TTS; free trial credit.
• Azure TTS, Amazon Polly, Google TTS — incumbent cloud options.
• Coqui TTS, Piper, F5-TTS — open-source, self-hostable.

Pick this if…

• Default production STT: Deepgram.
• Best accuracy + diarization: AssemblyAI.
• Self-host on a Mac / GPU: whisper.cpp / Faster Whisper.
• In-browser, no server: Transformers.js or Web Speech API.
• Production TTS: ElevenLabs (quality) or OpenAI TTS (cheap).
• Self-host TTS: Piper or F5-TTS.