AssemblyAI Review 2026: Speech-to-Text API, Pricing, Limits

Introduction

AssemblyAI sells speech-to-text through an API. It handles pre-recorded audio and realtime streams in 99 languages, and it adds speech understanding features on top of the transcript. In 2026 it also covers more of the voice stack, with a Voice Agent API, a Dictation API, guardrails and an LLM gateway. If you are building something that listens to audio, you could plausibly get most of what you need from this one vendor.

This review covers what AssemblyAI's own pricing page and documentation say, along with one independent comparison. scored.tools did not test the API hands-on. AssemblyAI scores 7.3 out of 10 overall. Integration is its strongest area. Accuracy evidence and pricing value are its weakest.

Key Features

Three ways to transcribe

AssemblyAI offers pre-recorded (async), realtime (streaming over WebSocket) and sync speech-to-text APIs. On the async side, the pricing page lists two models: Universal-3.5 Pro at the top and Universal-2 as the cheaper option. Streaming has its own models: Universal-3.6 Pro Realtime, plus Universal-Streaming in English and Multilingual versions. Choosing between them comes down to how much you care about accuracy versus cost per hour, which is a sensible setup for a transcription vendor.

Voice Agent API and Dictation API

The Voice Agent API bundles speech-to-text, the LLM, text-to-speech and turn detection into one priced product. That puts AssemblyAI up against orchestration platforms such as Vapi and Retell AI, and against voice vendors like ElevenLabs that have moved into agents. The Dictation API is aimed at apps where users talk instead of type. Having both next to the core transcription APIs is the main reason to consider AssemblyAI over a transcription-only provider.

Speaker diarization and PII redaction

Diarization (who spoke when) works on async and streaming audio. Redaction covers both the text and the audio itself, so you can strip personal data from the transcript and also bleep it out of the file. Profanity filtering and content moderation sit in the same guardrails group. Teams handling call recordings or healthcare audio need these, and the pricing page says a HIPAA BAA comes with no premium.

Speech understanding add-ons

On top of the transcript, AssemblyAI offers speaker identification, translation (pre-recorded only), entity detection, sentiment analysis, key phrases, topic detection and custom formatting. Keyterms prompting lets you bias the model toward domain words such as product names or drug names. Medical Mode is a separate add-on. The pricing page marks Auto Chapters and Summarization as deprecated and limited to Universal-2, which suggests AssemblyAI wants you to send that work through the LLM gateway instead.

LLM gateway

The LLM gateway sends transcripts to large language models through AssemblyAI's API, billed per million input and output tokens at rates that vary by model. It saves you from wiring up a second vendor for summaries or extraction. The downside is one more line item to watch.

SDKs, self-hosting and regions

The docs list multiple SDKs and documented integrations, and AssemblyAI says it offers a self-hosted deployment for voice AI. An EU region is priced the same as US. For teams with data residency rules, these options often decide the shortlist before accuracy is even discussed.

Pricing Breakdown

AssemblyAI publishes its rates, which plenty of speech vendors still don't. New accounts get $50 in credits with no credit card, according to the pricing page. Free-tier streaming is capped at 5 new streams per minute, pay-as-you-go at 100, and higher concurrency needs an enterprise contract.

Add-ons stack on the base rate

The headline rate is only where the bill starts. Each feature adds a per-hour charge:

• Diarization: +$0.02 async standard, +$0.065 async experimental, +$0.12 streaming
• PII redaction: +$0.08 text, +$0.05 audio
• Content moderation: +$0.15; profanity filtering: +$0.01
• Entity detection: +$0.08; topic detection: +$0.15; sentiment: +$0.02
• Medical Mode: +$0.15; keyterms prompting: +$0.05 async
• Voice Focus (Realtime Pro only): +$0.10

Here is how that adds up, using only the published numbers. A call-centre pipeline on Universal-3.5 Pro with standard diarization, text PII redaction and sentiment costs $0.21 + $0.02 + $0.08 + $0.02 = $0.33 per hour. That is 57% above the base rate. At 10,000 hours a month the gap is $3,300 versus $2,100. Multichannel audio is billed per channel, so a two-channel stereo call recording doubles the figure.

Streaming billing has a catch

Streaming is billed per WebSocket session hour, not per hour of audio. If your app opens a socket and leaves it idle while a user thinks, you pay for that silence. Close sessions aggressively, or the realtime bill will run ahead of your audio volume.

Volume pricing

The tool data lists custom plans from $0.05/hr, but the public pricing page says nothing about volume discount tiers. If you expect high volume, get a quote in writing before you commit an architecture to it.

Pros & Cons

Pros

• One API covers async transcription, streaming, dictation, voice agents, guardrails and LLM calls.
• Rates are published per hour, add-ons included, so you can model costs before talking to sales.
• $50 in signup credits with no card, plus API reference, guides and cookbooks.
• A self-hosted option and an EU region at US pricing.
• HIPAA BAA at no extra cost, per the pricing page.

Cons

• Accuracy on noisy audio is contested. A Speechmatics comparison page, based on G2 data, rates AssemblyAI at 80% against Speechmatics at 90% on noisy audio. Speechmatics is a competitor, so treat it as one data point, but it is the only independent figure in the evidence.
• AssemblyAI's claim of the lowest word error rate on the Pipecat benchmark comes from AssemblyAI itself.
• Add-ons stack quickly, and a full-featured pipeline can cost well above the headline rate.
• Streaming bills by session time, which penalises idle connections.
• Pay-as-you-go streaming tops out at 100 new streams per minute without an enterprise deal.
• Two summarisation features are already deprecated, a reminder that the feature list moves.

Who Is It For

AssemblyAI fits a developer or small product team that wants transcription and the layers around it from one vendor. If your roadmap runs from "transcribe uploads" to "redact PII" to "add a voice agent", it lets you stay on one SDK and one bill. The documentation scores well on ease of use, so getting a first request working should not take long.

It also suits regulated workloads that need a BAA, EU hosting or self-hosting, since those options are listed without extra fees for the first two.

It is a weaker pick if your audio is mostly noisy: field recordings, drive-thru orders, crowded call floors. The only independent comparison available points away from AssemblyAI there. Run your own sample set through AssemblyAI, Speechmatics and Deepgram before choosing. Pure voice-agent builders who care more about orchestration than transcription should compare the $4.50/hr Voice Agent API against Vapi and Retell AI, where you can pick each component yourself.

Verdict

AssemblyAI earns a 7.3/10. Its API coverage is the strongest in this category's evidence (9/10), and the docs make it easy to start. The weaker scores come from accuracy evidence that rests on the vendor's own benchmark against an independent comparison that rates it lower on noisy audio, and from add-on pricing that makes real bills hard to read from the headline numbers.

Recommendation: use AssemblyAI if you want one vendor for transcription, speech understanding and voice agents, and your audio is reasonably clean. Spend the $50 signup credit on a sample of your own worst recordings, price your exact add-on stack per hour, and get a volume quote if you expect more than a few thousand hours a month. If noisy-audio accuracy is the metric that decides your product, benchmark Speechmatics and Deepgram before you commit.