AssemblyAI

Our full take

AssemblyAI does one half of a voice agent, and it does it well. Think of an agent as having ears and a mouth. The ears turn what the caller says into text, the mouth turns the reply back into speech. AssemblyAI is the ears. It is speech-to-text plus what it calls speech understanding, and it does not speak back or place phone calls. So this is a building block, not a finished agent. Knowing that up front saves you comparing it like-for-like against an all-in platform.

The headline reason to reach for it is accuracy on the words that actually trip agents up. Names, email addresses, postcodes, order numbers. Its Universal-Streaming model is built for live agents and returns text in about 300 milliseconds (the launch post quotes a P50 of 303ms), which is fast enough that the caller does not feel the lag. AssemblyAI publishes that number itself, so treat it as a vendor claim until our own tests land, but the company’s reputation has long been built on transcription quality rather than marketing.

Here is the pricing, with the workings shown. AssemblyAI bills by the hour of audio, not per call, which is unusual in this directory and easy to misread. The real-time Universal-Streaming rate is $0.15 an hour. Divide by 60 and that is $0.0025 a minute, which is the figure on this page. That is genuinely cheap for the listening layer. For context, the cheaper async (pre-recorded) Universal-2 model is also $0.15/hr, the higher-accuracy Universal-3 Pro is $0.21/hr async, and the premium Universal-3 Pro streaming model runs $0.45/hr ($0.0075 a minute), which is the top of the range shown here. New accounts get $50 in free credit, no card needed, which is plenty to test with.

The “speech understanding” extras are where AssemblyAI separates from a plain transcriber, and they stack on the base rate per hour. Speaker identification is +$0.02/hr, sentiment +$0.02/hr, translation +$0.06/hr, entity detection +$0.08/hr, redacting personal data from the text +$0.08/hr, topic detection +$0.15/hr, and a medical mode for clinical terms +$0.15/hr. You only pay for the ones you switch on. For an agent that needs to redact a credit-card number on the fly or flag an angry caller, having those built into the same API is a real saving over wiring up three separate services.

One number to read carefully. AssemblyAI also lists a bundled Voice Agent API at $4.50/hr ($0.075/min) that wraps speech-to-text, a language model and a voice into one rate. That is a different product from the standalone transcription this profile rates, and the all-in math here deliberately covers only the speech-to-text, because that is the part AssemblyAI is known for and the part you would slot into a Vapi or LiveKit stack.

On languages, read the small print. The cheap Multilingual Universal-Streaming covers six languages, English, Spanish, French, German, Italian and Portuguese, all at the same $0.15/hr. The pricier Universal-3 Pro streaming model claims 99+ languages, so global reach exists, you just pay the premium streaming rate for it. If your callers are in those six big languages, the cheap tier is all you need.

Where AssemblyAI is unusually strong is the compliance paperwork, which matters if you handle health or payment data. Its own security page states SOC 2 Type 1 and Type 2, a GDPR third-party assessment, PCI-DSS 4.0 Level 1, AES-256 encryption at rest and TLS 1.3 in transit, with a choice of US or EU data residency. Separately, its docs confirm it will sign a HIPAA Business Associate Agreement (the contract that lets you process patient data legally), arranged through sales. That is more written assurance than several flashier rivals can show, and it is why this page ticks HIPAA where some others stay blank.

The honest limits. This is one piece of the puzzle, not the puzzle. No voice, no phone line, no built-in language model, so you are assembling the rest yourself or letting a platform like Vapi do it for you. It plugs neatly into Vapi, LiveKit, Daily and Pipecat-style stacks (that is how you give it a mouth and a phone number), but if you want to plug in one thing and have a working agent by lunchtime, AssemblyAI is not that thing. There is also no affiliate route to speak of: AssemblyAI’s own FAQ says plainly it does not run an affiliate programme, so there is no commission angle here, just the product.

My read. If the make-or-break for your agent is hearing the caller correctly, especially names, numbers and regulated data, AssemblyAI earns the shortlist at a low per-minute rate with the certificates to back a healthcare or finance build. If you want speech, dialling and orchestration handled in one bill, this is the wrong layer to start from, and a bundled platform will serve you better.

The 1 to 10 scores on this page are an editorial preview, our provisional read to get the framework in place, not a measured result. We have not run AssemblyAI through our own call tests yet, so there is no Voxrater latency figure here. The pricing, capability and compliance detail is sourced from AssemblyAI’s pricing, product, security and documentation pages, captured 2026-05-31.