Teams look past AssemblyAI when the invoice stops matching the headline rate, when a model generation they built on is retired, or when audio has to be processed on hardware they control. The twelve services in this category differ most on how the final hourly bill is assembled, where recognition can run, and whether a small vendor can be started without a sales call.
Short answer: The first three alternatives to AssemblyAI in our speech-to-text apis ranking are Soniox, Speechmatics and Deepgram. Challengers take places 1 to 3 here. Big brands follow in the order of our category ranking. Choose Soniox if you want one hourly figure with diarization already inside it. All 11 alternatives below are ranked products from Best Speech-to-Text APIs in 2026, and each says where it differs from AssemblyAI.
Vendors can pay for more visibility on this website, for example a higher place in a ranking or more room on a page. Commercial relationships can also affect which products we feature and how prominently. Every vendor link is marked the same, whatever the arrangement behind it. How that works.
Pricing: Published per audio hour by model; starter credit at signup; enterprise volume deals by contract
Established in: Not stated by the vendor
Built for: Developers who want to start on credits and pick analysis features one by one
Not for: Teams that need one flat all-in hourly price
01
Why buyers look past AssemblyAI
01
The base rate is only part of the hourly bill
AssemblyAI's pricing page lists pre-recorded audio at $0.21 per hour on AssemblyAI-STT-3.7 and $0.15 on Universal-2, with streaming from $0.30. Speaker identification, translation, entity detection, topic detection, medical mode and PII redaction are each priced per hour on top. A pipeline that switches on several of them pays a multiple of the advertised figure, so forecasting means adding up a list.
02
Model generations change underneath your integration
The pricing page marks Universal-3.5 Pro as deprecated and the streaming models Universal-Streaming English and Multilingual as replaced, and the language page says requests to retired models route to assemblyai-stt-3-7 unless you act. The new model covers 32 languages, against 99 on Universal-2, with automatic fallback rules that differ depending on whether language detection is switched on.
03
Deployment is a hosted API
The pages we read describe a cloud API with a signup credit of $50 and EU data residency at the same price as the US. Terms are with AssemblyAI Inc. at an address in New York, and the language page says nothing about self-hosting. Organisations that must run recognition inside their own network, or want a country other than the United States named on the contract, need to compare other options.
you hold a Microsoft agreement and need disconnected containers
#12
03
What each one does differently
1
Soniox
Not stated by the vendor · Published
Soniox quotes one price per audio hour, lower for recorded files than for live streaming, and folds speaker detection, language identification and smart formatting into that rate. That removes the add-on arithmetic described above. Live calls include translation at no extra cost. The pricing page mentions no starter credit, and only a cloud API is described.
Switch from AssemblyAI if you want one hourly figure with diarization already inside it.
Speechmatics prices by model tier, with a cheaper pre-recorded Melia 1 rate rising to Enhanced, and sells subscriptions and twelve-month credit packs. A $100 starting credit needs no card. A medical model exists, and containers and on-device deployment are available, though only inside the Enterprise tier. Single sign-on is a paid add-on.
Switch from AssemblyAI if you need containers or an on-premises option and accept an Enterprise contract.
Deepgram bills Nova-3 by the minute with no minimum spend, so small projects can start on pay-as-you-go with a $200 credit that carries no card requirement or expiry. Teams with steady volume can prepay Growth credits from $4,000 a year. Self-hosted containers on NVIDIA GPUs are reserved for Enterprise, and the home page names no headquarters.
Switch from AssemblyAI if you want a larger free credit and an upgrade path to self-hosted GPUs.
Rev AI charges Reverb transcription per hour and Whisper Large per minute, and the same pricing page lists a human transcription product at a per-minute price. New accounts get 5 hours of Reverb credit, and deployment can be cloud or on-prem. Forced alignment is offered. The footer reads only a copyright to Rev, so the ownership relationship stays unclear.
Switch from AssemblyAI if you sometimes need human-checked transcripts from the same vendor.
Voxist, a Paris company by its footer, sells speech recognition, synthesis and translation through an API next to contact-centre and voicemail products. Hosting ranges from a European cloud to a customer VPC, a data centre or an air-gapped site. Prices are quoted in euros, yet the pricing link opens a coming-soon page, and two certifications are described as in progress.
Switch from AssemblyAI if French-language audio must stay on European or fully isolated infrastructure.
Vosk is an Apache-2.0 toolkit rather than a billed service: recognition runs offline on a Raspberry Pi, Android, iOS or your own server, with per-language models of about 50 MB across 20+ languages. Bindings exist for Python, Java and C#. No usage fee appears, but you supply hosting, scaling and updates, and no company country is stated.
Switch from AssemblyAI if audio may never leave your own hardware and you can run the servers.
ElevenLabs sells monthly plans, from a free tier through Starter, Creator, Pro, Scale and Business to Enterprise, all drawing on one shared credit pool. Speech to Text uses 330 credits per minute, so 10k credits cover roughly half an hour. Scribe is missing from the pricing page, which leaves you converting credits to hours yourself.
Switch from AssemblyAI if you already pay for ElevenLabs voice products and want one credit balance.
Whisper's code and weights are MIT-licensed, and the README shows a pip install with local model loading, so you may run it yourself. OpenAI's hosted route lists Whisper at $0.006 per minute, gpt-4o-transcribe at the same rate and a mini variant at half. The pricing page shows no credits, and no add-on list inflates the rate.
Switch from AssemblyAI if you want a plain per-minute price, or a permissive licence to run models locally.
Google meters per minute and keeps two API generations with different rates. In v1, enabling data logging lowers the price, and v2 standard drops from $0.016 to $0.004 per minute at two million minutes monthly. The first 60 minutes each month are free, medical models cost several times more, and discounts come through a sales quote.
Switch from AssemblyAI if your volume is high enough to reach the lowest tier and you already use Google Cloud.
Amazon Transcribe counts audio per second in one-second steps with no minimum, at $0.006 per minute batch and $0.01 streaming in US East, varying by region. Diarization, language identification and custom vocabularies are inside the base rate. PII redaction, Call Analytics and summarisation cost extra, and a 12-month free tier gives 60 minutes monthly.
Switch from AssemblyAI if your workloads already run on AWS and short clips make per-second billing worthwhile.
Azure Speech bills per audio hour, counted per second, across Standard real-time, Fast Transcription, Batch, LLM Speech and Custom models. The F0 free tier gives 5 real-time hours a month without batch. Commitment tiers begin at 2,000 hours monthly, and containers run connected or disconnected. The price fields did not render on the page we read.
Switch from AssemblyAI if you hold a Microsoft agreement and need disconnected containers.
You want diarization and formatting inside one hourly price
Soniox: Speaker detection, language identification and smart formatting sit inside the per-hour rate, and live calls include translation.
02
You need recognition to run on your own servers or devices
Vosk: It is an Apache-2.0 toolkit with roughly 50 MB models for 20+ languages that works offline on Raspberry Pi, phones and servers.
03
Your French audio must stay on European or isolated hosting
Voxist: The Paris company lists hosting from a European cloud to an air-gapped site, though its pricing page shows no amounts yet.
04
You want the largest credit to test with before paying
Deepgram: Pay As You Go carries a $200 credit with no card and no expiry, and Nova-3 is billed per minute with no minimum.
05
Your short clips make per-second billing matter
Amazon Transcribe: Audio is counted in one-second steps with no minimum, and diarization is included in the base rate.
04
When AssemblyAI is still the right choice
AssemblyAI remains a reasonable choice when you want a documented API with a $50 signup credit, no card, and speech features such as speaker labels, summaries and redaction available behind the same key. EU data residency is offered at the same price as US processing. If you only use the base transcript and one or two add-ons, the bill stays predictable and a move gains little.
05
What moving off AssemblyAI involves
01
Re-measure accuracy on your own recordings
Word error rates differ by language, accent and audio quality, and AssemblyAI's own pages group Universal-2 languages into accuracy tiers. Run the same 50 to 100 representative files through each candidate and compare transcripts, speaker labels and timestamps before any contract.
02
Map the add-ons before you compare rates
List every add-on your code calls today, such as diarization, entity detection, summarisation or redaction, and check whether each candidate bundles it, sells it separately or does not offer it. Compare the all-in hourly cost for your actual feature set rather than the base figure.
03
Plan for response formats and streaming limits
Transcript JSON, word timestamps and webhook behaviour are not shared between vendors, so adapters need rewriting. For live audio, confirm concurrency limits, since AssemblyAI lists 5 new streams per minute on the free tier and 100 on pay-as-you-go.
06
Questions about replacing AssemblyAI
6 answers
Why does my AssemblyAI bill exceed the listed hourly rate?
The listed rate covers transcription only. Add-ons such as speaker identification at $0.02, translation at $0.06, entity detection at $0.08 and topic detection at $0.15 per hour are charged on top, and medical mode adds $0.15.
Which alternatives publish a free starting balance?
Deepgram lists a $200 credit with no card, Speechmatics $100 with no card, Rev AI 5 hours of Reverb credit, and AssemblyAI $50. Google and Amazon give monthly or 12-month free minutes instead.
Can any of these run without sending audio to a vendor cloud?
Vosk and Whisper run on your own hardware under open licences. Deepgram, Speechmatics, Rev AI, Voxist and Azure describe containers or on-prem options, mostly under enterprise terms.
Does one alternative include speaker detection in its base price?
Soniox folds speaker detection, language identification and smart formatting into its hourly rate, and Amazon Transcribe includes diarization in its base rate. AssemblyAI charges $0.02 per hour for async diarization.
Is there a service with its country stated as European?
Of the twelve, only Voxist states a country on the pages we read, France. Most others do not give one, so check the contracting entity in the terms before relying on jurisdiction.
How many languages does AssemblyAI cover?
Its language page lists 32 languages for AssemblyAI-STT-3.7 and 99 for Universal-2, with fallback to Universal-2 when a language is unsupported by the newer model.
—
Sources
Checked 18 November 2026
AssemblyAI pricing: Per-hour rates, add-on prices, $50 signup credit, concurrency limits, EU residency at same price, deprecated models.
Founder of The Knowledge Engineers and of 2 Steps Ahead, an AI search visibility agency. Editorially responsible for every page on this site; research and drafting are AI-assisted.
Author and fact-checker: Sebastiaan Smits. This page was last reviewed on 18 November 2026; our review process is public.