Mubert

Whisper is OpenAI's open-source, MIT-licensed speech-to-text model trained on 680,000 hours of audio -- you can download it and run transcription fully offline and free on your own hardware. It supports ~99 languages plus translation to English and is remarkably robust to accents and noise. If you don't want to run GPUs, OpenAI's hosted transcription API runs Whisper (and newer gpt-4o-transcribe models) at roughly $0.006/min. It has no built-in speaker diarization and the core repo updates infrequently, but the surrounding ecosystem (whisper.cpp, faster-whisper, WhisperX) is enormous.

Cartesia builds real-time-first voice models -- its Sonic TTS and Ink STT rank #1 on Artificial Analysis speech leaderboards for combined quality and speed. Built on state-space (Mamba-style) architectures for ultra-low latency, it's purpose-made for voice agents and powers platforms like Retell. One developer API covers TTS, STT, and voice agents, with a genuinely usable free tier (20K credits/mo) and paid plans from $5/mo, plus cloud, on-prem, and on-device deployment. The main friction is an abstract credit model and promo pricing that muddies the long-term cost.

AssemblyAI is a developer-first speech-to-text platform built for production voice AI, with industry-leading accuracy from its Universal model family. Beyond transcription it offers a Speech Understanding API (summarization, sentiment, PII redaction), a Voice Agent API, and an LLM Gateway (formerly LeMUR) that runs LLMs over transcripts at passthrough pricing. Pricing is transparent pay-as-you-go -- pre-recorded from $0.15/hr, realtime from $0.15/hr -- with an unusually generous free tier (185 hours pre-recorded, 333 hours streaming, no card). It's a builder's tool, not a consumer app, and competes head-to-head with Deepgram.

Industry-leading AI voice platform for text-to-speech, voice cloning, and multilingual dubbing. Produces the most natural-sounding synthetic speech with instant cloning from short samples. Used by podcasters, game devs, and SaaS companies building voice features. Robust API for easy integration.

Deepgram is a developer speech platform best known for fast, cheap, accurate speech-to-text via its Nova model family, plus Aura text-to-speech and a voice-agent API. Pricing is pay-as-you-go per minute (Nova STT from roughly $0.0077/min, with promotional rates lower) and $200 in free credits to start, making it one of the cheapest production STT options. It's optimized for real-time, high-throughput voice applications and competes directly with AssemblyAI. Like AssemblyAI, it's infrastructure for builders, not a consumer-facing tool.

Stable Audio is Stability AI's music and sound-effects generator, and the only major player offering open-weight music models trained on fully licensed data. The hosted app (running Stable Audio 2.5) has tiers from free to $89.99/mo, while the Stable Audio 3.0 Small and Medium models released in May 2026 are open weights on Hugging Face, free for commercial use under $1M revenue. That means you can self-host, own your outputs, and generate variable-length tracks up to six minutes. The hosted free tier is thin (10 generations, 30-second crop, non-commercial), but the open-weight option is genuinely unique.