Choosing a speech to text API used to mean picking the least bad Whisper wrapper. In 2026 the field is crowded with real options, prices have fallen to cents per hour, and the differences that matter are no longer obvious from a headline accuracy number. This guide compares twelve APIs on what you actually pay, whether they stream, how many languages they cover, and how to judge their accuracy claims, with every price taken from the vendor's own pricing page.
The short answer: for developer and technical speech, Avalon (Aqua Voice's API) is the strongest pick and works with the OpenAI SDK, at $0.39 per audio hour. For real-time voice agents, Deepgram and AssemblyAI lead. Soniox is the cheapest at about $0.10 per hour. Google, AWS, and Azure make sense if you already live on their cloud.
How we compared
Every price below is the pay-as-you-go list price for the vendor's current flagship model, read on the vendor's own pricing page on October 28, 2026. We converted per-minute and per-token prices to per audio hour so the numbers line up; the vendor's original unit is noted where it differs. Volume discounts, committed-use plans, and enterprise contracts can bring every one of these lower, so treat the table as a starting point rather than a quote.
We graded each API on one real reason to pick it, and we were careful with accuracy. Vendors publish their own word error rate (WER) numbers on their own test sets, which are rarely comparable. The fairest public yardstick is the Hugging Face Open ASR Leaderboard, and even that changes its dataset mix over time. More on reading benchmarks below.
Quick comparison
API | Batch price (per audio hour) | Streaming price (per audio hour) | Languages (vendor's wording) | Free to start |
|---|---|---|---|---|
Avalon (Aqua Voice) | $0.39 | Batch only | Any ISO 639-1 hint or auto-detect | First $1.00 free |
OpenAI | $0.27 ($0.0045/min) | $1.02 ($0.017/min) | Not stated for gpt-transcribe | None stated |
Deepgram | $0.26 ($0.0043/min) | $0.29 promo, $0.46 regular | 45+ | $200 credit |
AssemblyAI | $0.21 | $0.15 to $0.45 | 18 (3.5 Pro), 99 (Universal-2) | Free tier, no card |
ElevenLabs | $0.22 | $0.39 | 90+ | 4.5 hours included |
Soniox | about $0.10 | about $0.12 | 60+ | None stated |
Speechmatics | $0.24 base | See vendor | 56 | $100 credit |
Rev AI | $0.20 (English) | Streaming API available | English (Reverb) | 5 hours of credit |
Gladia | €0.16 to €0.27 | €0.68 | 6 (Solaria-3), 100+ (Solaria-1) | €50 credit |
Google Cloud | $0.18 (dynamic batch) | $0.96 | Not stated | $300 Google Cloud credit |
Amazon Transcribe | $0.36 ($0.006/min) | $0.60 ($0.01/min) | Not stated | 60 min/month for 12 months |
Azure AI Speech | $0.18 (batch) | $1.00 | Not stated | 5 hours/month (real-time) |
Prices at the time of writing, October 28, 2026. Azure's pricing page did not render its numbers, so Azure figures come from Microsoft's published Azure Retail Prices list.
The best speech to text APIs in 2026
1. Avalon (Aqua Voice): best for developer and technical speech
Avalon is the speech model behind the Aqua Voice dictation app, sold as an API. It was trained on how people actually talk to computers: prompting a coding agent, reading out a terminal command, naming a model version. That is where general-purpose models stumble, turning "zshrc" into "C sharp C" or "GPT-4o" into "GPT-4.0." If your audio is full of product names, code, or AI jargon, this is the reason to pick it.
The independent proof: when Avalon debuted on the Hugging Face Open ASR Leaderboard in October 2025, it was the #1 proprietary model, with a 6.24% average WER and #6 overall across open and proprietary models. We wrote that up in Avalon on the OpenASR Leaderboard. The leaderboard has since changed its dataset mix, so that placement is a dated result, not a current rank. On AISpeak, our own self-reported benchmark of real-world clips full of AI terms, Avalon got the key term right more often than every model we tested, including ElevenLabs Scribe and Whisper Large v3; the full results are in Introducing Avalon.
The practical reason developers switch is the interface. Avalon serves the OpenAI request, response, and error shapes, so the official OpenAI SDKs work after you change two lines:
from openai import OpenAI
client = OpenAI(api_key="your-avalon-api-key", base_url="https://api.aquavoice.com/v1")
transcript = client.audio.transcriptions.create(model="avalon-v1.5", file=open("speech.mp3", "rb"))
print(transcript.text)
What you get, per the API documentation: ten audio formats up to 256 MiB and four hours per file, speaker labels for up to 8 speakers, a prompt field for vocabulary hints, segment timestamps, SRT and VTT output, and durable async jobs you can poll. Pricing is $0.39 per audio hour, billed per second with a 10-second minimum, no seat fees, and the first $1.00 of usage free with no card.
Honest limits: Avalon is batch only, with no streaming endpoint, so it is not the pick for a live voice agent. It is not the cheapest per hour on this list. And its benchmark story is strongest in English; for many languages at once, look at ElevenLabs, Soniox, or Speechmatics. The same platform also offers personalized dictation routes and an MCP server if you are building voice into an app rather than transcribing files.
2. OpenAI: best if you are already on the OpenAI platform
OpenAI's current recommendation is gpt-transcribe for recorded audio, at $0.0045 per minute (about $0.27 per hour), and gpt-live-transcribe for live audio, at $0.017 per minute (about $1.02 per hour). Both accept prompt, keywords, and languages hints.
The news that matters for anyone with an existing integration: OpenAI's speech to text guide states that whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize "are deprecated and will shut down on February 26, 2027." Word timestamps and diarization lived on those older models, and OpenAI notes that gpt-live-transcribe does not return word-level timestamps or speaker labels. If you have to touch your transcription code anyway, it is a good moment to test alternatives. Because Avalon keeps OpenAI's request shapes, it is one of the few you can try without rewriting the client; our Aqua vs OpenAI Whisper page covers the differences.
3. Deepgram: best for real-time voice agents
Deepgram is built around streaming. Its Flux models are built for streaming voice agents, and Nova-3 covers general transcription with monolingual and multilingual versions plus medical and pharma variants. Nova-3 pre-recorded is $0.0043 per minute (about $0.26 per hour). Streaming Nova-3 is $0.0048 per minute on what Deepgram labels a "limited-time promotional" rate, against a regular $0.0077; Flux English is $0.0065 per minute. Keyterm prompting, redaction, and streaming diarization are paid add-ons.
You start with a $200 credit and no card. Deepgram says its Nova models support 45+ languages. The reason to pick it: a streaming-first platform with models made for voice agents.
4. AssemblyAI: best balance of price and features
AssemblyAI prices by the hour and keeps those prices low. Universal-3.5 Pro (pre-recorded) is $0.21 per hour, the older Universal-2 is $0.15 per hour, and streaming runs from $0.15 per hour (Universal-Streaming) to $0.45 per hour for the new Universal-3.6 Pro Realtime. Diarization, keyterm prompting, and a new Medical Mode are add-ons priced per hour.
Its free tier needs no card, and AssemblyAI says it covers up to 185 hours of pre-recorded or 333 hours of streaming transcription. Watch the language counts by model: Universal-3.5 Pro "works across 18 languages," Universal-2 supports 99, and the cheapest streaming model is English only. The reason to pick it: a deep feature set at some of the lowest hourly prices from a major vendor.
5. ElevenLabs Scribe: best for many languages and speakers
Scribe v2 costs $0.22 per hour for batch and Scribe v2 Realtime $0.39 per hour, the same rate on every ElevenLabs plan. ElevenLabs states 90+ languages, speaker diarization for up to 32 speakers, word-level timestamps, keyterm prompting for up to 1,000 terms, and a Scribe v2 Medical model at the batch price. The realtime model is listed at about 150 ms latency. The free plan includes about four and a half hours of Scribe v2. The reason to pick it: broad language coverage and strong diarization at a low batch price.
6. Soniox: lowest price per hour
Soniox prices by token, which works out, in its own words, to "about $0.10/hour for async (file) and $0.12/hour for real-time (streaming) transcription," the lowest list price on this page. It covers 60+ languages in one model, supports diarization, timestamps, and custom terms, and includes translation at no extra cost. Soniox's pricing page shows no free tier. The reason to pick it: high-volume transcription where cost per hour decides.
7. Speechmatics: best for mixed-language audio
Speechmatics' Melia 1 model "transcribes 56 languages without choosing one up front," which matters for calls and meetings where speakers switch languages mid-sentence. Read its pricing page carefully: the headline $0.12 per hour for Melia 1 is a discounted figure, and the page itself explains that it starts from a $0.24 base rate before training-program and annual-commitment discounts. Oak 1 is its medical model. New accounts get $100 in credit with no card. The reason to pick it: code-switching audio across many languages.
8. Google Cloud, Amazon Transcribe, and Azure AI Speech: best if you already live on that cloud
The three hyperscalers have the broadest locale tables and deep enterprise features (custom vocabularies, custom language models, PII redaction, medical products), and billing that goes on the cloud account you already have.
- Google Cloud Speech-to-Text runs Chirp 3 on its V2 API. Standard recognition is $0.016 per minute (about $0.96 per hour) for the first 500,000 minutes a month, while dynamic batch is $0.003 per minute (about $0.18 per hour) if you can wait. New Google Cloud accounts get $300 in credits.
- Amazon Transcribe charges $0.006 per minute for batch and $0.01 per minute for streaming (about $0.36 and $0.60 per hour), billed per second with no minimum, and includes 60 free minutes a month for 12 months. Transcribe Medical and Call Analytics are priced separately.
- Azure AI Speech lists real-time transcription at $1.00 per hour and batch at $0.18 per hour in Microsoft's retail price list, with 5 free real-time hours a month on the F0 tier. Microsoft documents diarization for up to 35 speakers and custom speech models.
The reason to pick one: procurement, data residency, and integration with the rest of your stack, rather than price or raw accuracy.
Also worth a look
- Rev AI sells its Reverb English model at $0.20 per hour, plus a separate foreign-language model, a streaming API, and human transcription. New accounts get five hours of credit.
- Gladia prices in euros: Solaria-3 is €0.16 per hour for batch but covers 6 languages, Solaria-1 is €0.27 per hour across 100+ languages, and real-time is €0.68 per hour. New accounts get €50 in credit for a year.
How to read speech to text benchmarks
Accuracy claims are the hardest part of this market to compare, so a few rules of thumb:
- Word error rate depends on the test set. A vendor's "5% WER" on its own data and another's "8% WER" on different data say nothing about each other. Prefer results where every model ran on the same audio.
- Public leaderboards are a snapshot. The Open ASR Leaderboard is the most widely cited neutral comparison, but it adds and removes datasets over time, so a rank from one year is not comparable with the next. Always ask which date and which datasets.
- Average WER hides the words you care about. A model can score well overall and still mangle every product name, drug name, or command in your audio. If your domain has its own vocabulary, measure key-term accuracy, which is why we built AISpeak.
- Test on your own audio. Take 30 to 60 minutes of real recordings, have them transcribed by hand, and run each candidate on the same files. Many vendors on this list give you enough free credit to do this for nothing.
How to choose in one minute
- Your audio is full of code, product names, or AI terms: Avalon.
- You are migrating off whisper-1 or gpt-4o-transcribe and want the smallest code change: Avalon (same SDK) or OpenAI's gpt-transcribe.
- You are building a live voice agent: Deepgram or AssemblyAI streaming.
- You transcribe thousands of hours and price decides: Soniox or AssemblyAI Universal-2.
- Your audio mixes languages or needs 90+ of them: Speechmatics or ElevenLabs Scribe.
- Your company already runs on Google Cloud, AWS, or Azure: start with that cloud's service.
Not building an app? You may want dictation instead
Many people searching for a speech to text API really want to talk instead of type on their own computer. You do not need an API key for that. Aqua Voice is a dictation app for Mac, Windows, iPhone, and Android that runs on Avalon, types into any app, keeps a custom dictionary for your terms, and lets you edit text by voice. It is free for 1,000 words, then $10/mo billed annually or $12/mo month to month for Pro. Aqua has a SOC 2 Type II audit report and is ISO 27001 certified, and is GDPR and HIPAA compliant. Our best dictation software roundup compares it with the alternatives, the medical dictation roundup covers clinical use, and the user guide shows setup.
If you are building, create an API key, spend the free dollar on your own audio, and compare.
FAQ
What is the best speech to text API in 2026?
It depends on your audio. For developer and technical speech, Avalon from Aqua Voice is the strongest pick and works with the OpenAI SDK at $0.39 per audio hour. For real-time voice agents, Deepgram and AssemblyAI lead. Soniox has the lowest list price, at about $0.10 per hour. Google, AWS, and Azure suit teams already on those clouds.
What is the cheapest speech to text API?
At list prices on October 28, 2026, Soniox is the cheapest at about $0.10 per hour for files and $0.12 for streaming, followed by AssemblyAI Universal-2 and Universal-Streaming at $0.15 per hour and Gladia's Solaria-3 at €0.16 per hour. Volume and committed-use discounts change the ranking at scale, so price your real monthly hours.
Is the OpenAI Whisper API being shut down?
OpenAI's documentation says whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize are deprecated and will shut down on February 26, 2027, with gpt-transcribe and gpt-live-transcribe recommended instead. Avalon keeps OpenAI's request and response shapes, so you can test it as a replacement by changing the base URL and model name.
Does Avalon support real-time streaming?
No. Avalon is a batch transcription API: you send a file (up to four hours) and get the transcript back, either synchronously or as an async job. For live streaming, choose a streaming API such as Deepgram, AssemblyAI, ElevenLabs Scribe v2 Realtime, or OpenAI's gpt-live-transcribe.
