Speech to Text API: the 2026 Field, Priced and Compared

A speech-to-text API is a web service that converts audio into text programmatically: you upload a file or stream audio frames over a WebSocket, and the service returns a transcript (often with timestamps, speaker labels, and confidence scores) that your application can process. The 2026 market spans batch/file APIs (AssemblyAI Universal, OpenAI transcription models, Speechmatics Melia), realtime streaming APIs (Deepgram Nova-3, ElevenLabs Scribe v2 Realtime, AssemblyAI Universal-Streaming), and voice-agent bundles (Deepgram Voice Agent API, AssemblyAI Voice Agent + LLM Gateway). Typical billing is per minute or per hour of audio, with separate rates for batch and realtime.

Effective rates run from about \$0.10 to \$1.44 per audio hour depending on provider and mode. The floor as of September 2026: Microsoft MAI-Transcribe-2 at \$0.10/hr (early-bird through 2026, public preview), AssemblyAI Universal-2 at \$0.15/hr, OpenAI gpt-4o-mini-transcribe at \$0.003/min (\$0.18/hr), Deepgram Nova-3 batch at ~\$0.26/hr, ElevenLabs Scribe v2 at ~\$0.22/hr, Speechmatics Melia 1 at \$0.129/hr. OpenAI's newly recommended GPT Transcribe runs \$0.0045/min - 50% above the mini model it displaced from the top of the pricing page. Streaming and feature add-ons (diarization, redaction, keyterms) can double any of these effective rates.

Groq is the only major provider with a genuinely free speech-to-text tier: whisper-large-v3 and whisper-large-v3-turbo at 2,000 requests/day, 20 requests/minute, no credit card, no expiry - roughly 8 hours of audio per day with chunking, 25 MB per request. Everyone else offers one-time signup credits instead: Deepgram \$200, AssemblyAI \$50, Speechmatics \$100 (Speechmatics replaced its recurring 3,000-minute monthly free allowance with the one-time credit on July 31, 2026). Google Cloud gives a recurring but tiny 60 free minutes/month. Beyond credits, the genuinely free path is local: Whisper's MIT-licensed weights, whisper.cpp, faster-whisper, and NVIDIA Parakeet run on your own hardware at \$0 per minute.

On independent leaderboards, ElevenLabs Scribe v2 leads commercial ASR at 2.2% WER (Artificial Analysis, after the 2026 price cut to ~\$3.67 per 1,000 minutes), with Microsoft's MAI-Transcribe-1.5/2 close behind (2.0-2.6%) and AssemblyAI, Deepgram Nova-3, and OpenAI's GPT-4o transcribe models clustered within 1-3 points of open Whisper Large-v3. NVIDIA's open Canary-Qwen 2.5B tops the Hugging Face Open ASR Leaderboard at 5.63% WER. Accuracy on paper matters less than correction time: accented, crosstalk, or domain-specific audio shifts rankings more than a fraction of a WER point, so test candidates on your own audio before committing.

Use a cloud API when you need sub-300 ms realtime streaming for voice agents, an SLA, or built-in audio intelligence (LLM-on-transcript, PII redaction, sentiment in one call). Go local when privacy matters (audio never leaves the machine), volume is steady (self-hosting crosses Deepgram-class API rates around 300-500 hours/month), or you want a flat cost with no meter. Local Whisper-class engines now sit within 1-3% WER of the cloud APIs. VocalFuse is the turnkey local route on Windows: Whisper-class engine, hold-to-talk dictation, file transcription, bot-free meeting capture, Basic \$5/mo flat with no per-minute billing; DIY builders get the same engine free via whisper.cpp or faster-whisper.

For non-urgent batch, Google Cloud STT v2 Dynamic Batch is \$0.003/min (\$0.18/hr) for up-to-24-hour turnaround, and OpenAI gpt-4o-mini-transcribe is also \$0.003/min with immediate results. Microsoft's MAI-Transcribe-2 undercuts both at \$0.10/hr - but read the caveats: it is public preview with no SLA and the rate is promotional through end of 2026 with no published permanent price. At real scale, discounts compound: Speechmatics auto-applies 20% above 500 hrs/month, Google's volume tiers reach \$0.004/min past 2 million minutes, and self-hosting wins outright past a few hundred steady hours. Quote your full pipeline (diarization, redaction, storage) - add-ons can double the sticker.

Microsoft MAI-Transcribe-2 ships diarization, word-level timestamps, keyword biasing, and code-switching inside the base \$0.10/hr rate - the bundle that pressures specialist vendors who meter those features separately. AssemblyAI includes speaker labels in its per-hour rates. Deepgram charges separately for diarization (\$0.0020/min), redaction, keyterm prompting, and entity detection - the effective rate for a fully featured pipeline can exceed twice the base model price. ElevenLabs Scribe v2 includes diarization and posts the highest published speaker-label accuracy (~98%). OpenAI charges a premium for the diarize variant (\$0.015/min vs \$0.006 standard).

Yes. Two paths: (1) self-host the actual engine - Whisper's MIT-licensed weights run via whisper.cpp or faster-whisper, and wrapper projects expose OpenAI-compatible endpoints (for example, whisper-local serves POST /v1/audio/transcriptions on localhost:7777, so existing API calls work unchanged with the cloud swapped out); (2) use a turnkey app - VocalFuse packages the same Whisper-class engine as a Windows desktop product with dictation, file transcription, and meeting capture for \$5/mo flat, no API code required. Both keep audio on your machine; the DIY path is free, the turnkey path trades \$5/mo for not owning the ops.