AssemblyAI

Name: AssemblyAI
Brand: AssemblyAI
Price: 0.15250 USD

The speech and understanding API behind other people's conversation intelligence

AssemblyAI is a speech-to-text and speech-understanding API that developers call to transcribe recorded or streaming audio and then extract conversation intelligence signals from it, including speaker labels, sentiment, topic and entity detection, auto chapters, key phrases, summarization, PII redaction, and LLM-generated analysis; it is priced pay-as-you-go from $0.15 per hour of transcription with $50 in free credits and self-serve signup, and it ships no user interface, dashboard, or CRM integration of its own.

Visit website

Overview

Say the important thing first: AssemblyAI is not an app. There is no recording bot, no call library, no manager dashboard, no scorecard, no talk-ratio report, no deal board, and no CRM connector. If you are a sales manager who wants to coach five reps next Monday, this is the wrong product and nothing in it will help you. AssemblyAI is the layer underneath products like that, and a meaningful number of the conversation intelligence tools a small business evaluates are calling an API like this one behind the scenes.

What it sells is speech-to-text plus a set of understanding models on top. You post an audio file or open a streaming connection, and you get back a transcript with speaker labels, plus optional analysis: sentiment per utterance, topics classified against the IAB taxonomy, more than fifty entity types, automatic chapter segmentation, key phrases, a summary, redacted PII in both transcript and audio, and content moderation flags. An LLM Gateway lets you run frontier models over the transcript for anything the fixed models do not cover, which is how most teams build call scoring and custom trackers.

The commercial shape is genuinely small-business friendly, which is why it belongs on this list despite having no interface. Signup is self-serve with $50 in free credits and no credit card, pricing is published to four decimal places, and there is no minimum commitment or annual contract. Universal-2 transcription runs $0.15 per hour and the newer Universal-3.5 Pro $0.21 per hour, with understanding features stacking additively in small increments: speaker diarization at $0.02 per hour, sentiment at $0.02, entity detection at $0.08, translation at $0.06. A startup transcribing a thousand hours a month with speaker labels and sentiment is looking at under two hundred dollars.

The company is a serious one. Founded in 2017 by Dylan Fox, headquartered in San Francisco, roughly a hundred employees, and about $158M raised across rounds including a $50M Series C led by Accel in December 2023 with Y Combinator, Insight Partners, Nat Friedman, and Daniel Gross among backers. It is one of the two credible independent speech API vendors alongside Deepgram, and the practical choice between them comes down to price per hour, latency profile, and how much of the analysis you want the vendor to do for you.

Best for

Software teams building a product that needs transcription and conversation analysis inside it, agencies with an unusual analysis requirement no packaged tool covers, and technically capable small companies that would rather own their call data and pay by the hour than pay per seat.

Not the right fit for

Sales managers who want to coach reps; there is no interface, no scorecard, no library, and no dashboard, so buying this instead of a packaged tool means writing an application first.
Anyone who needs CRM write-back, deal boards, pipeline risk, or forecasting, because AssemblyAI has no concept of a deal, an account, or an opportunity.
Teams with no engineering capacity. The free credits and published prices make it look approachable, but every useful output requires code you maintain forever.
Buyers who want a fixed monthly bill; usage pricing means the cost tracks call volume, and a busy month costs more than a quiet one with no cap unless you build one.
Real-time products built by teams that have not read the billing model, since streaming charges on connection time rather than speech time and idle sessions accumulate charges silently.

How it works

1
For recorded audio you upload a file or point the API at a URL and submit an async transcription job, choosing a model and toggling the understanding features you want. The response is a JSON transcript with word-level timings, and each enabled feature adds its own structured block: speaker labels, sentiment per utterance, detected entities, chapter boundaries, key phrases, or a summary.
2
For live audio you open a streaming websocket connection and receive partial and final transcripts as the person speaks. Streaming is billed on session duration rather than speech duration, meaning an open but idle connection still costs money, which is the single most common billing surprise for teams building real-time products on top of it.
3
Conversation intelligence gets built by combining the primitives. Speaker diarization plus word timings gives you talk-to-listen ratio and longest monologue with arithmetic you write yourself. Entity and key phrase detection plus your own term list gives you competitor and keyword tracking. Sentiment gives you the tone curve across a call. None of it arrives as a chart; it arrives as JSON for you to store and render.
4
For anything judgement-shaped, such as scoring a call against a MEDDICC rubric or deciding whether the rep confirmed a next step, you send the transcript through the LLM Gateway to a frontier model with your own prompt, billed per million tokens at the provider's rate. That is how teams build the scorecard layer that a packaged tool would sell them, and it is both cheaper per call and considerably more work.

Feature breakdown

25 features in 4 modules

Transcription models

The base layer, priced per hour with no seats involved.

Universal-2 async transcription: Pre-recorded transcription at $0.15 per hour, the workhorse model for batch processing a call archive where latency does not matter.
Universal-3.5 Pro async transcription: The higher-accuracy pre-recorded model at $0.21 per hour, worth the six cents on audio with accents, crosstalk, or technical vocabulary.
Universal streaming transcription: Real-time streaming from $0.15 per hour of session time, for products that need words on screen while the person is still speaking.
Universal-3.5 Pro Realtime: The premium streaming tier at $0.45 per hour base, with inline speaker labels available in the stream rather than only after the fact.
Medical mode: A domain-tuned option adding $0.15 per hour across models for clinical vocabulary, relevant if your conversations are healthcare rather than sales.
Word-level timestamps: Every word carries a timing, which is the raw material for talk ratio, monologue length, interruption detection, and clip generation in whatever you build.

Speech understanding models

The conversation intelligence primitives, each priced as a small per-hour add-on.

Speaker identification and diarization: Who spoke when, at $0.02 per hour on async and $0.12 per hour on streaming. Without this nothing downstream can distinguish the rep from the prospect.
Sentiment analysis: Sentiment classification per utterance at $0.02 per hour, which is how you build a tone curve across a call or across an account relationship.
Topic detection: Classification against the IAB taxonomy so calls can be grouped by subject without you writing a classifier.
Entity detection: More than fifty entity types at $0.08 per hour, covering people, organizations, dates, monetary amounts, and locations, which is the substrate for competitor and pricing trackers.
Auto chapters: Automatic segmentation of a long call into chapters with summaries, the feature that makes a ninety-minute recording navigable.
Key phrases: Extraction of the salient phrases in a conversation, useful as a cheap first pass before spending LLM tokens.
Summarization: Model-generated summaries of a transcript without a separate LLM call, adequate for a recap and not for a scorecard.
Translation: Transcript translation at $0.06 per hour, for teams whose customer calls are not all in one language.
Custom formatting: Structured output shaping so the transcript arrives in the form your application expects rather than requiring post-processing.

Privacy, safety, and guardrails

The compliance primitives, which matter more in this category than most buyers expect.

PII redaction in transcript: Personally identifiable information can be stripped from the text before it lands in your database, which is how you record customer calls without storing card numbers.
PII redaction in audio: Redaction applied to the audio itself, not only the transcript, which is the version that actually satisfies a PCI or healthcare conversation.
Content moderation: Flagging of sensitive content at up to $0.15 per hour, the most expensive guardrail and the one regulated industries need.
Profanity filtering: The cheapest guardrail at $0.01 per hour, for products that surface transcripts to customers.
You control retention: Because the transcript lands in your own storage, retention policy is a decision you make rather than a tier you buy, which is the single biggest structural advantage of the API route.

LLM Gateway and voice agents

Where custom scoring and interactive products get built.

LLM Gateway: Run frontier models over your transcripts through the same account and bill, priced per million tokens at the underlying provider rate, with Claude and GPT models available.
Custom call scoring via prompts: Scorecards, playbook adherence, and objection classification are built by prompting a model over the transcript, which gives you exactly the rubric you want and none of the interface a packaged tool would provide.
Voice Agent API: An all-inclusive $4.50 per hour endpoint combining transcription, reasoning, text to speech, and tool calling, for teams building an agent that talks rather than one that listens.
Concurrency limits by tier: Free accounts are limited to 5 new streams per minute against 100 for pay-as-you-go, which is the practical gate between prototyping and production.
SDKs and documentation: Client libraries and a documentation set widely regarded as among the best in the speech API market, which is a real cost saving on a small engineering team.

Use cases

4 documented

SaaS team adding call analysis to their own product

Customers are asking for meeting summaries and searchable transcripts inside the application, and building a speech stack in-house is a year of work.

Transcription with diarization and summarization is wired in over a sprint, billed per hour of customer audio rather than per seat, so the feature's cost scales with usage instead of headcount.

Agency with an analysis requirement no tool covers

A client needs calls scored against a compliance script with a bespoke rubric, and every packaged conversation intelligence tool enforces its own scorecard model.

Transcripts go through the LLM Gateway with a custom prompt, producing exactly the rubric the client specified, at a per-call cost measured in cents rather than a per-seat subscription.

Startup that wants to own its call data

Two years of customer calls locked inside a vendor's library, with retention limits set by the plan tier and no clean export path.

Recording and storage stay in the company's own infrastructure with AssemblyAI used only for processing, so retention, residency, and deletion are internal policy decisions rather than a pricing tier.

Support operation building automated QA

Only two percent of support calls get reviewed, and a QA platform priced per agent is not affordable at the volume the team actually runs.

Every call is transcribed and scored automatically with sentiment, PII redaction, and prompt-based rubric scoring, at a cost driven by minutes of audio rather than the number of agents.

Pricing

from $0.15 per hour of pre-recorded transcription (Universal-2), with $50 in free credits

Usage-based pay-as-you-go priced per hour of audio for transcription, with each speech understanding feature added as a separate per-hour increment and LLM usage billed per million tokens. No seats, no minimum commitment, no annual contract.

Plan	Price	Includes
Free credits	$50 credit one time on signup	No credit card required Full access to transcription and understanding models Limited to 5 new streams per minute for streaming Roughly 330 hours of Universal-2 transcription at list price
Pay-as-you-go transcription	$0.15 to $0.21 per hour of pre-recorded audio	Universal-2 at $0.15 per hour Universal-3.5 Pro at $0.21 per hour Whisper-class accuracy without hosting your own models No monthly minimum
Pay-as-you-go streaming	$0.15 to $0.45 per hour of session time	Universal streaming from $0.15 per hour Universal-3.5 Pro Realtime at $0.45 per hour base 100 new streams per minute concurrency Billed on connection duration, not speech duration The session-duration billing rule is the one thing to model carefully before building on streaming.
Speech understanding add-ons	$0.01 to $0.15 per hour, per feature, additive	Speaker diarization $0.02 async, $0.12 streaming Sentiment analysis $0.02 Entity detection $0.08 Translation $0.06 Profanity filtering $0.01, content moderation up to $0.15
Voice Agent API and LLM Gateway	$4.50 per hour (agent); per-token for LLMs usage	Voice Agent at $4.50 per hour all-inclusive LLM Gateway billed at underlying provider token rates Claude and GPT model families available Single account and single bill across speech and LLM usage

Billing notes

Understanding features stack additively on top of the base transcription rate, so a call processed with diarization, sentiment, and entity detection costs the base rate plus twelve cents an hour, not a bundled premium.
Streaming is billed on session duration from connection open to close, which means idle time and silence are charged; this is the most common source of unexpected bills.
Free tier accounts face a streaming concurrency limit of 5 new streams per minute against 100 on pay-as-you-go, which is a production gate rather than a volume gate.
There is no seat concept at all, so the per-recorded-user versus per-team-member question that dominates the rest of this category simply does not apply.
Volume commitments and enterprise agreements exist but are not required; a small company can run in production on published pay-as-you-go rates indefinitely.
Because you store the transcripts, there are no retention tiers and no storage charges from the vendor, which shifts both the cost and the compliance responsibility to you.

Value assessment: On raw arithmetic the API route is dramatically cheaper than seats. A five-rep team on 25 hours of calls each per month is 125 hours, which with Universal-2 plus diarization and sentiment costs under twenty-five dollars, against several hundred dollars a month for a packaged tool. That comparison is also dishonest unless you price the engineering. You are buying JSON, and everything a manager would actually use, the library, the search, the scorecard, the dashboard, the CRM sync, is software you build and maintain. AssemblyAI is excellent value when you are building a product or have an unmet analysis requirement, and terrible value as a way to save money on a tool you could just buy.

Strengths & limitations

Strengths

Transparent published pricing to four decimal places with no seats, no minimum, and no annual contract, which is rare in a category built on opaque enterprise quotes.
The understanding models cover most conversation intelligence primitives out of the box: diarization, sentiment, topics, entities, chapters, key phrases, and summarization, so you are not building classifiers from scratch.
PII redaction applied to audio as well as transcript, which is the version that satisfies regulated conversations and which many packaged tools do not offer at all.
The LLM Gateway means custom scoring rubrics and bespoke trackers are a prompt rather than a feature request to a vendor who will say no.
Documentation and SDK quality is among the best in the speech market, which materially reduces the engineering cost of the route.
You keep the data. Recording storage, retention windows, deletion, and residency become your own policy rather than a plan tier, which is the strongest argument for the API approach.
A stable, well-funded independent vendor: founded 2017, roughly 100 employees, about $158M raised including a $50M Series C led by Accel.

Limitations

There is no product for an end user. No library, no dashboard, no scorecards, no coaching workflow, no deal view, and no CRM integration of any kind.
Every useful output requires engineering, and the ongoing maintenance of that code is a cost that never appears in the per-hour price comparison.
Streaming billed on session duration rather than speech duration punishes naive implementations and idle connections.
Usage pricing means no cost ceiling by default; a spike in call volume produces a spike in the bill unless you build your own limits.
Speaker diarization on streaming costs six times the async rate at $0.12 per hour, which changes the economics of real-time coaching products specifically.
Nothing in the product addresses recording consent, two-party consent law, or disclosure; those obligations sit entirely with whatever application you build on top.

Head-to-head comparisons

5 alternatives

AssemblyAI vs Deepgram

from $0.0043 per minute (Nova-3 pre-recorded monolingual), with $200 in free credits

Deepgram is the direct competitor and prices lower on the base model, roughly $0.0043 per minute for Nova-3 pre-recorded against AssemblyAI's $0.15 per hour, with $200 in free credits and diarization included rather than charged. AssemblyAI counters with a broader speech understanding set, audio-level PII redaction, and an LLM Gateway in the same account. Choose Deepgram for lowest cost per minute and latency-critical streaming; choose AssemblyAI when you want the analysis models and the LLM layer supplied rather than assembled.

Full AssemblyAI vs Deepgram comparison

AssemblyAI vs Symbl.ai

from $0 free tier, then $0.027 per minute of audio or video pay-as-you-go

Symbl.ai sits higher up the stack, offering conversation-native APIs such as a Call Score endpoint priced per criteria per conversation and a real-time assist API, now under Invoca's ownership. AssemblyAI gives you cleaner, cheaper primitives and expects you to build the scoring. Take Symbl if you want scoring as an API call; take AssemblyAI if you want control and better unit economics on volume.

Full AssemblyAI vs Symbl.ai comparison

AssemblyAI vs Speak AI

from $0 monthly with $2.00 per hour transcription (Pay as you go), or $20 per user per month (Pro)

Speak AI is roughly thirteen times the price per hour at $2.00, and for that premium you get an actual application: capture, storage, custom scoring against your own methodology, thematic analysis, and delivery, with no code involved. AssemblyAI gives you the primitives at $0.15 an hour and nothing else. If you have engineers, AssemblyAI plus your own application wins on cost and control; if you do not, Speak AI is what you would have spent six months building.

Full AssemblyAI vs Speak AI comparison

AssemblyAI vs Grain

from $0 (Free, capped at 20 meetings), then about $15 per seat per month on Starter billed annually

Grain is the packaged product most buyers in this category actually want: recording, a shared library, coaching, and CRM sync at around $29 a seat with free viewer licenses and nothing to build. AssemblyAI is the layer somebody would use to build a Grain. If you have a sales team and no engineers, this comparison is not close and Grain wins outright; if you are shipping call analysis inside your own software, only the API makes sense.

Full AssemblyAI vs Grain comparison

AssemblyAI vs Avoma

from $19 per seat per month (Startup, billed annually)

Avoma is the full meeting lifecycle platform with scorecards, trackers, dialer capture, and a revenue intelligence module, from $19 a seat plus add-ons. AssemblyAI has none of that and will never have it. The only reason to compare them is if you are deciding whether to buy conversation intelligence or build it, and the honest answer for most small businesses is buy.

Full AssemblyAI vs Avoma comparison

Implementation & onboarding

Setup time: An hour to a first transcript: sign up, take the API key, post an audio file, read the JSON. Weeks to months to anything a non-engineer would recognize as a product, because storage, search, playback, and reporting are all yours to build.
Learning curve: Low for a developer, thanks to unusually good documentation and SDKs. Infinite for a non-developer, since there is no interface to learn.
Onboarding: Fully self-serve with $50 in free credits and no credit card. Moving from free to pay-as-you-go raises streaming concurrency from 5 to 100 new streams per minute, which is the main production gate.
Migration notes: Switching between speech APIs is genuinely easy compared with switching packaged tools, since the interface is a transcription request and a JSON response. The hard part of any migration here is your own accumulated post-processing, not the vendor. Historical audio can be reprocessed in bulk at the same per-hour rate.

Platform, API & security

Platforms: REST APIStreaming websocket APIPython, JavaScript, and other client SDKsLLM Gateway
API: The product is the API. Async transcription, real-time streaming, speech understanding models, LLM Gateway, and a Voice Agent endpoint, all self-serve with published per-hour and per-token pricing.
Compliance: SOC 2GDPRHIPAA-oriented handling available via medical mode and PII redaction
Data residency: Enterprise arrangements exist; not a self-serve selectable option on pay-as-you-go.
SSO: Account-level authentication with API keys; SSO is an enterprise agreement concern rather than a published tier feature.
Security notes: PII redaction is available in both transcript and audio, and content moderation and profanity filtering are separately priced guardrails. Because transcripts are returned to you rather than stored in a vendor library, retention, deletion, and residency are governed by your own infrastructure. Recording consent and two-party consent obligations are entirely the responsibility of the application built on top.

Support & resources

Channels: Email supportDeveloper support on paid usageEnterprise support agreements
Documentation: Extensive developer documentation at assemblyai.com covering every model, parameter, and understanding feature, widely rated among the best in the speech API market.
Community: Active developer community through Discord, GitHub examples, and a substantial technical blog and tutorial library.

Company

Founded: 2017
Headquarters: San Francisco, California, United States
Ownership: Venture-backed
Founders: Dylan Fox
Employees: Approximately 101 (reported 2026)
Funding: Approximately $158M raised, including a $50M Series C in December 2023 led by Accel with Insight Partners, Y Combinator, Keith Block, Smith Point Capital, Nat Friedman, and Daniel Gross participating.

Funding history

Round	Amount	Year	Notes
Seed and Series A	Part of the $158M total	2019 to 2021	Early backing from Y Combinator and Accel.
Series B	$28M	2022	Led by Insight Partners as the speech-to-text market consolidated around a few independent vendors.
Series C	$50M	2023	Led by Accel in December 2023, bringing total funding to roughly $158M.

Timeline

2017Founded by Dylan Fox in San Francisco to sell speech-to-text as a developer API rather than as an enterprise platform.
2019Goes through Y Combinator and establishes the pay-as-you-go, published-pricing model that still defines the product.
2022Raises a $28M Series B led by Insight Partners and expands beyond transcription into speech understanding models.
2023Raises a $50M Series C led by Accel in December, reaching roughly $158M total funding.
2024Ships the Universal model family and an LLM layer so developers can run frontier models over transcripts without leaving the platform.
2026Universal-3.5 Pro arrives for async and realtime, alongside an LLM Gateway with Claude and GPT models and a $4.50 per hour Voice Agent API.

Integrations

REST API for pre-recorded audio
Streaming websocket API for real-time audio
Python, JavaScript, and community client SDKs
LLM Gateway to Claude and GPT model families
Any application, CRM, or data warehouse you connect it to yourself
No native meeting platform, CRM, or dashboard integrations

Frequently asked questions

11 questions

What is AssemblyAI?

AssemblyAI is a speech-to-text and speech-understanding API. Developers send it recorded or streaming audio and get back a transcript with speaker labels, plus optional analysis including sentiment, topics, entities, auto chapters, key phrases, summaries, PII redaction, and content moderation. It is infrastructure, not an application.

Does AssemblyAI record my sales calls or give me a dashboard?

No. There is no recording bot, no call library, no search interface, no scorecard, no talk-ratio report, no deal board, and no CRM integration. Everything a manager would use has to be built by your engineers on top of the API. If you want a tool you can log into on Monday and coach reps with, buy Colibri, Spiky, Avoma, or Grain instead.

How much does AssemblyAI cost?

Pre-recorded transcription is $0.15 per hour on Universal-2 and $0.21 on Universal-3.5 Pro. Streaming starts at $0.15 per hour and the premium realtime model is $0.45 per hour base. Understanding features add small per-hour amounts: speaker diarization $0.02, sentiment $0.02, entity detection $0.08, translation $0.06. Signup includes $50 in free credits with no credit card.

Is there a per-seat price?

No, and this is the structural difference from the rest of the category. There are no seats at all, so the question of whether you pay per recorded user or per team member does not arise. You pay for hours of audio processed, which means a hundred people can view the results without changing the bill.

Can I build call scoring and coaching metrics on top of it?

Yes, and this is the normal pattern. Speaker diarization plus word timings gives you talk-to-listen ratio and longest monologue with your own arithmetic. Entity and key phrase detection gives you competitor and keyword tracking. For judgement-based scoring against a rubric such as MEDDICC, you send the transcript through the LLM Gateway with your own prompt. All of it is code you write and maintain.

How does streaming billing work and why do people get surprised?

Streaming is charged on session duration, meaning from when the websocket connection opens to when it closes, not on how much speech was actually transmitted. An implementation that opens connections early, holds them through silence, or fails to close them cleanly accumulates charges for nothing. Model this before building a real-time product.

What does AssemblyAI do about recording consent and two-party consent law?

Nothing, by design. It processes audio you send it. Obtaining consent, announcing recording, and complying with two-party consent jurisdictions are obligations of the application you build. What it does provide is PII redaction in both transcript and audio, and content moderation, which are the compliance primitives you would use to keep sensitive data out of your own store.

Where are transcripts stored and how long are they retained?

In your own infrastructure. AssemblyAI returns the transcript to you and you decide where it lives, how long it is kept, and when it is deleted. There are no retention tiers and no storage charges, which is a real advantage over packaged tools where history limits are a pricing lever, and a real responsibility because compliance becomes yours.

AssemblyAI or Deepgram?

Deepgram prices lower on the base model, includes speaker diarization on pre-recorded audio rather than charging for it, and gives $200 in free credits against AssemblyAI's $50. AssemblyAI offers a broader understanding model set including auto chapters and audio-level PII redaction, plus an LLM Gateway in the same account. For pure transcription volume Deepgram usually wins on cost; for analysis supplied rather than built, AssemblyAI usually wins.

Is AssemblyAI a stable company to build on?

Reasonably. It was founded in 2017 by Dylan Fox, is headquartered in San Francisco with around a hundred employees, and has raised roughly $158M including a $50M Series C led by Accel in December 2023. It is one of the two credible independent speech API vendors, and switching between speech APIs is far easier than switching packaged tools, which limits the downside of the bet.

What languages does it support?

The Universal model family covers a broad multilingual set with dedicated multilingual streaming and async models, and transcript translation is available as a $0.06 per hour add-on. Exact language coverage differs by model, so check the current model documentation rather than assuming parity between the async and streaming paths.

Editorial verdict

AssemblyAI is the best-documented, most transparently priced way to put speech understanding into software you are building, and it deserves its place on this list only for buyers who are building. The understanding models cover the conversation intelligence primitives properly, PII redaction reaches the audio itself, the LLM Gateway makes custom scoring a prompt rather than a roadmap request, and the per-hour economics beat per-seat pricing by an order of magnitude on paper. The caveat is not subtle: there is no product here. No dashboard, no library, no scorecards, no CRM sync, and nothing at all for a sales manager. If you have engineers and an unmet requirement, this is an excellent choice. If you have a sales team and a coaching problem, buy a packaged tool and do not let the per-hour price tempt you into a build.

Written by the SaaSTracker editorial team. Awards, when shown, are judged against the published criteria in our methodology.