10 Best Speech-to-Text APIs of 2026

AssemblyAI splits its API into pre-recorded and streaming models and prices both by the hour. For file transcription, Universal-2 covers 99 languages at $0.15 per hour, while Universal-3.5 Pro costs $0.21 per hour and focuses on 18 languages with code switching and stronger diarization. For live audio, Universal-Streaming costs the same $0.15 per hour for English, a multilingual version adds Spanish, French, German, Italian and Portuguese at the same price, and Universal-3.6 Pro Realtime costs $0.45 per hour for 32 languages with about 150 ms median latency, context carryover and voice isolation. Streams return partial and final transcripts with endpointing and word timestamps, and AssemblyAI lists unlimited concurrent streams. The catch is the add-on menu: keyterms prompting, diarization, medical mode and speaker focus each carry a per-hour fee on some models. Python and JavaScript SDKs keep the WebSocket setup short.

Deepgram offers two main speech models. Nova-3 handles general transcription in 60+ languages with diarization, smart formatting and keyterm prompting, and Flux STT targets voice agents with turn detection and interruption handling. Streaming prices on pay as you go are $0.0077 per minute for Nova-3 monolingual or Flux English and $0.0092 for Nova-3 multilingual, which works out to $0.46 and $0.55 per hour; Deepgram says pre-recorded audio costs less. Billing is per second without block rounding, and smart formatting is included, while redaction costs $0.0020 per minute and keyterm prompting $0.0013. The $200 starting credit needs no card and does not expire. Deployment is the main differentiator: you can use Deepgram's cloud, route through regional endpoints, or self-host, and you can train a custom model on your own data. The Growth plan cuts rates by up to 20% but starts at $4,000 a year.

Speechmatics prices each model separately and publishes every rate. For pre-recorded audio, Melia 1 costs $0.12 per hour and transcribes 56 languages without being told which one to expect, switching mid-sentence when speakers do. The Enhanced model costs $0.38 per hour, adds a custom dictionary and is the vendor's most accurate option, and Oak 1 handles clinical language for $0.15 per hour in batch. For live audio, Linden 1 is built for voice agents with turn detection at $0.15 per hour, and Melia 1 real-time costs $0.20. Speechmatics says real-time transcripts arrive in under one second. New accounts get $100 in credits with no card, credit packs start at $25, and volume discounts reach 25%. The company also lists on-prem and on-device deployment, which few rivals offer at this price. Add-ons such as translation at $0.49 per hour and chapters at $0.30 add up quickly.

Scribe is the speech-to-text side of ElevenLabs, sold through the same API that runs its voices. Scribe v2 handles files in 90+ languages with keyterm prompting and dynamic audio tagging at $0.22 per hour, and Scribe v2 Realtime streams live audio with a published latency under 150 ms for $0.39 per hour. Keyterm prompting adds $0.05 per hour on batch and $0.08 on realtime, and transcript editing adds $0.07 or $0.12. A medical variant is HIPAA-eligible, with BAAs and a zero retention mode, and ElevenLabs says Scribe also supports on-premise configurations for enterprises. API usage is billed in US dollars per audio minute rather than in credits, which makes cost easy to forecast, and a free plan with monthly credits covers early tests. It also plugs directly into ElevenLabs Agents if you are building a voice agent there. The main drawback is price: realtime costs more than double AssemblyAI's streaming rate.

Gladia sells one speech-to-text layer with transcription and audio intelligence bundled into the same request. Its Solaria models cover 100+ languages in both batch and real-time modes and handle sentences that switch language mid-flow without breaking timestamps. Diarization, sentiment analysis, named entity recognition and custom vocabulary are part of the core API, and add-ons such as translation and PII redaction are billed per hour only when you turn them on. Gladia states GDPR, HIPAA, SOC 2 Type II and ISO 27001 compliance and offers EU data residency, which matters for European call recording. Self-serve plans include 30 concurrent requests, rates drop automatically as monthly volume grows, and large top-ups earn bonus credits. Pricing from our location was shown only in euros, so we could not confirm a USD rate; check the vendor's page from a US account before you budget.

OpenAI's speech-to-text sits inside the same API most AI developers already use. For recorded audio, gpt-transcribe costs $0.0045 per minute and accepts keyword lists, language hints and a free-text prompt describing the recording, which helps with product names and account codes. For live audio from a microphone or phone call, gpt-live-transcribe streams text as the speaker talks for $0.017 per minute. Diarization runs through gpt-4o-transcribe-diarize, which can match known speaker names, but OpenAI notes speaker labels are not supported in realtime sessions. The older gpt-4o-mini-transcribe remains available at $0.003 per minute, and the original whisper-1 model is deprecated, so new projects should start on gpt-transcribe. Uploads are capped at 25 MB, which means long recordings must be compressed or split. OpenAI does not publish a latency figure for live transcription. The setup is simple if you already have keys and billing in place.

Rev AI is the API arm of Rev, a long-running transcription service, and that history shows in its pricing. Reverb, its in-house model, transcribes English at $0.20 per hour and Spanish, French, Chinese, Portuguese and 53 more languages at $0.30 per hour, with files rounded up to the nearest second and a 15 second minimum. A Whisper Large option costs $0.005 per minute, and the same API can route a file to human transcribers for $1.99 per minute, which few rivals offer. Insight APIs cover language identification, translation, summarization, sentiment and topic extraction. Rev AI supports async and streaming transcription, SDKs, and cloud or on-prem deployment, and lists SOC II, HIPAA, GDPR and PCI compliance with 99.99% uptime. Free credits equal 5 hours of Reverb. The list is lighter on voice agent features such as turn detection, and Rev AI does not publish a streaming latency figure.

Azure AI Speech bundles speech to text, text to speech and translation in one Azure service. For transcription you can pick real-time, fast or batch modes, train a custom speech model on your own audio and vocabulary, and caption audio in more than 100 languages. Continuous language identification, diarization and pronunciation assessment are enhanced add-ons billed per feature. Embedded speech runs recognition on devices, and connected containers let you host models in your own environment under commitment tiers. Microsoft provides SDKs in C#, C++, Java and other languages. The free F0 tier includes 5 audio hours of standard real-time transcription a month and 5 hours of custom real-time transcription, which is generous for testing. Billing is per second. The pricing page loaded its rates with a script from our location and showed no USD figures, so we marked the price as pending. Expect more setup than a single-vendor API, since you create Azure resources and choose regions.

Google Cloud Speech-to-Text gives you a set of recognition models behind one API: Chirp, the default and latest models, and task-tuned options such as phone_call and video, plus medical conversation and dictation models. It supports 85+ languages and variants, speaker diarization, model adaptation for custom vocabulary, automatic punctuation and language detection. Standard recognition costs $0.016 per minute for the first 500,000 minutes a month and $0.01 per minute after that, with deeper discounts at higher tiers, and the first 60 minutes each month are free. Requests round up to the nearest second. New Google Cloud customers also get $300 in general credits. The trade-off is effort and price: you need a Cloud project, billing and service accounts, and the starting rate is several times what AssemblyAI or Speechmatics charge. Google does not publish a single streaming latency figure on the product page.

Amazon Transcribe is the default speech-to-text option inside AWS. It runs on what Amazon calls a multi-billion parameter speech foundation model for both streaming and recorded speech, and it offers features across 100+ languages, including automatic punctuation. You can train custom language models, add custom vocabulary, redact sensitive content automatically and flag toxic speech. The medical side is HIPAA-eligible and connects to AWS HealthScribe for clinical documentation. Billing is per second of audio with no minimum, and the free tier covers 60 audio minutes a month for the first 12 months, though not redaction or custom language models. Rates differ by AWS region and the page loaded them with a script, so we could not confirm a USD rate and marked pricing as pending. Setup requires an AWS account, IAM permissions and a region choice, which adds steps compared with a single-key API, but it keeps audio and transcripts under your existing AWS controls.