Best Speech-to-Text API 2026: 13 Transcription APIs Compared

As of July 2026, OpenAI's gpt-4o-mini-transcribe at $0.003/minute ($0.18/hour) is the cheapest list-price API from a major vendor. Soniox is effectively cheaper at ~$0.0017/minute on its token-based async pricing, and Rev AI Reverb matches OpenAI at $0.003/min. At very high volumes (>5M minutes/month) AWS Transcribe drops to $0.0078/min. There is no cheaper managed option without committing to volume tiers.

ElevenLabs Scribe v2 Realtime publishes 150 ms first-word latency on the FLEURS benchmark — the only number we found measured against a public dataset. Gladia Solaria-1 publishes sub-103 ms partial-token latency (a different metric). Soniox reports ~200 ms. Deepgram Nova-3 reports under 300 ms. For voice-agent use cases (LLM completion budget around 800 ms), any of these leaves headroom.

Google Cloud Speech-to-Text V2 with Chirp 2 lists 125+ languages — the broadest. Gladia Solaria-1 lists 100, including languages not offered by other commercial APIs (long-tail African, Central Asian, and Pacific). AssemblyAI Universal-2 and ElevenLabs Scribe v2 each support 99. Real-time language coverage is always smaller than batch — verify your specific languages are streaming-supported before architecting on real-time.

Not anymore. whisper-1 is now considered legacy; OpenAI itself recommends gpt-4o-transcribe for new projects (better accuracy, same $0.006/min) or gpt-4o-mini-transcribe for cost (half the price). Whisper Large-v3 open-source weights — which you self-host — score around 6.4% WER on the Hugging Face Open ASR Leaderboard. Strong, but a few open-source models now beat it on that benchmark.

Batch (also called async or prerecorded) accepts a complete audio file and returns the full transcript after processing — typically 5–30 seconds for short files, sub-real-time for long files. Real-time (also called streaming) processes audio chunks as they arrive, returning partial transcripts within 150–600 ms. Real-time costs 1.5–5× batch on most vendors. Choose batch unless your use case truly needs live captions.

Diarization (separating speakers without prior voice samples) is included on AssemblyAI Universal-2, Deepgram Nova-3, ElevenLabs Scribe v2, Google Cloud STT V2, AWS Transcribe, Azure AI Speech, Speechmatics, Gladia, Rev AI, and Soniox. OpenAI does not provide diarization on the transcription endpoint — you need a separate library (Pyannote, NVIDIA NeMo) post-hoc.

Yes. AWS Transcribe Medical, Azure AI Speech, Deepgram Nova-3, AssemblyAI, Rev AI (HIPAA Plus subscription), Google Cloud STT, and ElevenLabs Scribe v2 all sign Business Associate Agreements. OpenAI signs BAAs only on its Enterprise tier — standard API accounts cannot ship clinical workflows. Treat any vendor that won't sign a BAA as not HIPAA-eligible regardless of marketing language.

Chunk on silence boundaries, never on fixed time intervals. In Python: pydub.silence.split_on_silence(audio, min_silence_len=500, silence_thresh=-40). From the CLI: ffmpeg -af silencedetect=n=-30dB:d=0.5. Add 0.5-second overlap on each chunk and dedupe overlapping words at boundaries. Better: switch to a vendor without the 25 MB cap (Deepgram, AssemblyAI, Gladia all accept multi-GB uploads).

Yes. OpenAI Whisper is open-weights (MIT) since 2022. NVIDIA Canary and Parakeet variants are open-source and currently lead the Open ASR Leaderboard. Self-hosting costs roughly $0.001–$0.003/min on a reserved A10G/L4 GPU at 50–80% utilization, beating most APIs at scale, but adds MLOps overhead, eval pipelines, and on-call burden. For most teams, the developer-time-versus-cost tradeoff favors a managed API up to several thousand hours per month.

Both are OpenAI hosted endpoints. whisper-1 is the original Whisper Large model from 2022, $0.006/min, REST-only (no streaming). gpt-4o-transcribe (March 2025) is built on the GPT-4o family — better accuracy on noisy audio, supports streaming via the OpenAI Realtime API, accepts a prompt parameter for context priming, and costs the same $0.006/min. For new projects OpenAI recommends gpt-4o-transcribe. whisper-1 stays available for backward compatibility.

Every vendor in this guide has a free tier or free credit. Most generous: Deepgram ($200 credit, no card, ~26,000 batch minutes), AssemblyAI ($50 credit, ~333 hours of Universal-2), ElevenLabs (10 hrs/mo Scribe v2), Gladia (10 hrs/mo), Speechmatics (480 min/mo). Google Cloud and AWS offer 60 minutes/month for 12 months. OpenAI has no free tier — you pay from minute one.

Word Error Rate degrades 2–4× on noisy audio (background music, reverb, low SNR) versus clean studio audio. Open ASR Leaderboard numbers (5–15% WER on benchmark sets) become 15–35% on phone-quality recordings. Deepgram Nova-3 and AssemblyAI Universal-2 publish noise-robustness benchmarks; most vendors do not. Beyond WER, the Koenecke et al. FAccT 2024 paper documented that ~1% of Whisper outputs hallucinate (entirely fabricated phrases) on real-world audio — independent of WER. Run voice-activity detection pre-processing and evaluate hallucination rate separately on your own data before shipping.

It is a credible option for batch workloads. Speechmatics positions Melia-1 as its 2026 flagship, priced at $0.24/hr batch — matching the cheapest AssemblyAI tier — with 50+ languages and diarization included. The main constraint is that Melia-1 is batch-only, with no real-time streaming endpoint, so streaming products need Speechmatics Ursa-2 or another vendor. As with any vendor, run an evaluation on a sample of your own audio before committing.

Two reasons. First, vendors typically publish best-case results on cherry-picked datasets (usually LibriSpeech clean, an audiobook benchmark) — real-world audio is harder. Second, WER normalization matters: results computed with lowercasing and punctuation stripped are not comparable to results computed without it, and different benchmarks make different choices. Treat vendor-published English WER as a best case rather than a production estimate, and always benchmark on a sample of your own audio before committing.

As of August 2026 verified pricing: AssemblyAI Universal-2 at $0.15/hr ($0.0025/min) is the cheapest managed tier — dropped from $0.24/hr in July 2026. Next cheapest is OpenAI GPT-4o Mini Transcribe at $0.18/hr, followed by Speechmatics Melia-1 at $0.24/hr. Below list price: AWS Transcribe at $0.0078/min above 5M min/mo, and self-hosted Whisper on reserved GPU (~$0.001-$0.003/min at 50-80% utilization). For predictable-volume workloads, flat-rate bulk options often beat per-minute API pricing.

It is a reasonable candidate, especially if you are already on OpenAI. OpenAI positions gpt-4o-transcribe as the recommended successor to whisper-1, with better accuracy on noisy audio, streaming via the Realtime API, and a prompt parameter for context priming, at $0.006/min. Practical constraints: a 25 MB limit on the REST endpoint, and no Open ASR Leaderboard submission, so there is no neutral cross-vendor accuracy reference for it. Evaluate it on audio representative of your own workload — including the longest files you expect to handle — before committing.

VexaScribe is a consumer transcription product, not an API vendor. We do not sell a transcription API and we do not resell any of the vendors covered here. We use Speechmatics Melia-1 as our own primary engine with Whisper Large-v3 as fallback, and we disclose that. We have no affiliate relationships with any vendor listed and received no compensation, free credits, or other consideration for inclusion. Note that we published a first-party benchmark on this page in July 2026 and withdrew it in August 2026 after finding methodological problems in our own work; see the correction on this page.

It depends on your languages. Deepgram lists 36+ languages, which trails Google Cloud STT V2 with Chirp 2 (125+), Gladia Solaria-1 (100), and AssemblyAI Universal-2 and ElevenLabs Scribe v2 (99 each), so start by confirming your specific languages are supported — and, if you need streaming, that they are supported in real time. Deepgram's strengths are streaming latency (under 300 ms first-word) and call-center features. If those are not what your multilingual workload needs, evaluate a broader-coverage vendor alongside it on your own audio.