Speech-to-text APIs · Comparison
Deepgram vs AssemblyAI vs Speechmatics: AssemblyAI 8.4, Speechmatics 8.1, Deepgram 6.6; 500 hours cost $85 to $129
We read the pricing, docs, data-usage, trust and legal pages of three transcription APIs on the same day and scored them on the same six indicators. What 500 hours of diarized batch audio cost, how each handles your audio, where the BAA sits, and who should pick which.
On this page9 sections
On six indicators read from each vendor's public pages, AssemblyAI scores 8.4 out of 10, Speechmatics 8.1 and Deepgram 6.6. For our reference scenario, 500 hours a month of pre-recorded English with speaker labels and word timestamps, run as batch jobs, AssemblyAI Universal-2 comes to $85 a month, Speechmatics Melia 1 to $120 and Deepgram Nova-3 Monolingual to $129, all at pay-as-you-go list prices. AssemblyAI stands out for compliance controls you switch on yourself: a BAA signed online, a training opt-out and retention as short as an hour. Speechmatics stands out for its privacy defaults, with no training on your audio unless you opt in and batch files deleted after 7 days. Deepgram stands out for the largest free credit and speaker labels included in its batch rate.
AssemblyAI and Speechmatics are in the Excellent band and Deepgram in the Good band. The scores measure what the public pages commit to, not transcription accuracy or speed, which depend on your audio, and we did not sign up for any of them.
At a glance
| Deepgram | AssemblyAI | Speechmatics | |
|---|---|---|---|
| What you pay for | Per-minute rates with speaker labels included in batch | Per-hour models with every add-on priced, and self-serve compliance controls | Credits used at hourly rates, with training off by default |
| Free credit | $200, no card | $50, no card, does not expire | $100, no card, valid 12 months |
| Reference scenario (500 h batch with diarization) | $129 on Nova-3 Monolingual, computed | $85 on Universal-2, computed | $120 on Melia 1 at list price, computed |
| Streaming | $0.0077/min regular; $0.0048 limited-time promotional | From $0.15/h (Universal-Streaming) | From $0.40/h (Melia 1, list) |
| Streaming diarization | +$0.0020/min | +$0.12/h | Included on Enhanced, Standard and Linden |
| Training on your audio | By default; opt out per request | By default; paid accounts opt out online | Only if you opt in |
| HIPAA BAA | Enterprise, through sales | Signed online on paid plans, no cost | Sub-BAA through the Trust Center; process not stated |
| Regions | EU, Australia and India endpoints | EU, at the US price | EU, US and Australia for all customers |
| Score (public pages) | 6.6 Good | 8.4 Excellent | 8.1 Excellent |
What the public pages show
Deepgram bills each second of audio from prepaid credit. Nova-3 Monolingual costs $0.0043 a minute for pre-recorded audio with speaker diarization included, and streaming runs on limited-time promotional rates below the regular $0.0077 a minute. Keyterm prompting, redaction and streaming diarization are per-minute add-ons. Growth prepays $4,000 or more a year for up to 20% off, and Enterprise, priced by sales, adds the BAA, self-hosted containers and custom models.
AssemblyAI prices each model per audio hour: Universal-2 at $0.15 for 99 languages, Universal-3.5 Pro at $0.21 for 18, and streaming from $0.15 an hour of session time. Diarization, PII redaction, keyterm prompting, Medical Mode and translation each have an hourly price. Paid accounts get a self-serve BAA, a training opt-out and a retention setting in Data Controls.
Speechmatics sells credits that each model uses at a fixed hourly rate. Melia 1 lists at $0.24 an hour for pre-recorded audio and $0.40 for real-time, with diarization, custom dictionary and language identification included on the models that support them. Its pricing page leads with rates after a 33% model-training discount and a 25% annual commitment discount; we used the list rates.
Strengths and things to know
Deepgram. Strengths: speaker labels included in its $0.0043-a-minute batch rate; a $200 credit with no card; EU, Australia and India endpoints on the same keys; self-hosted containers on Enterprise and a SageMaker route. Things to know: requests are used for model improvement unless each one sets mip_opt_out=true; the BAA is for Enterprise customers and DPAs are entered into on request; streaming rates are promotional with no end date; and the pages disagree on the Growth discount and on whether credits expire.
AssemblyAI. Strengths: a BAA signed online at no cost, a training opt-out and a retention setting down to one hour on paid accounts; a DPA that applies with the terms; ISO 27001:2022 and SOC 2 Type 2; EU processing at the US price. Things to know: free accounts cannot opt out of training; streaming bills session time, idle included; and a few add-on prices differ between the pricing page, pricing.md and the changelog.
Speechmatics. Strengths: no training on your audio unless you opt in; batch files deleted after 7 days and real-time audio not stored; EU, US and Australia regions for all customers; on-prem, virtual appliance and on-device deployment on Enterprise. Things to know: the headline prices include optional discounts; PII redaction, SSO and on-prem are Enterprise features; Melia 1 lacks real-time speaker labels and custom dictionary for now; and the credit expires after 12 months.
Pricing reality
Reference scenario: 500 audio hours a month (30,000 minutes) of pre-recorded English, run as asynchronous batch jobs with speaker diarization and word-level timestamps, on the cheapest generally available model that does both, at pay-as-you-go list prices in the US before tax, without free credit, commitment or training discounts, or promotional rates.
| Service | Model | Monthly cost | Where the number comes from |
|---|---|---|---|
| AssemblyAI | Universal-2 | $85 | Computed: 500 × ($0.15 + $0.02 diarization) |
| Speechmatics | Melia 1 | $120 | Computed: 500 × $0.24 list rate; diarization included |
| Deepgram | Nova-3 Monolingual | $129 | Computed: 30,000 × $0.0043; diarization included |
Discounts can change the order. Speechmatics' model-training opt-in brings Melia 1 to $0.16 an hour, or $80 a month for the same work, in exchange for letting it train on your audio. Deepgram's Growth rate of $0.0036 a minute would make it $108 a month, but Growth needs a prepaid commitment of at least $4,000 a year, more than a year of this workload at that rate (our arithmetic).
Add-ons move the totals too. PII redaction adds $40 a month on AssemblyAI ($0.08 an hour) and $60 on Deepgram ($0.0020 a minute), and is an Enterprise feature at Speechmatics. Live audio costs more everywhere: with speaker labels, 500 hours of streaming would cost $291 on Deepgram at the regular rate, $135 on AssemblyAI Universal-Streaming (session time, idle included) and $225 on Speechmatics Standard, where speaker labels are included (our arithmetic).
How we scored them
Every product in this category is scored on the same six indicators, read from the vendor's own pricing, docs, data-usage, trust and legal pages, with fixed formulas. The overall score is the plain mean. Our methodology page explains how these scores work.
| Indicator | Deepgram | AssemblyAI | Speechmatics |
|---|---|---|---|
| Pricing clarity | 4 | 5 | 5 |
| Free credit | 10 | 9.7 | 9 |
| Price for 500 hours | 6.8 | 8.2 | 7.1 |
| Transcription features | 7.9 | 9.1 | 9 |
| Data use and retention | 4.5 | 8.5 | 10 |
| Compliance and deployment | 6.5 | 10 | 8.5 |
| Overall | 6.6 | 8.4 | 8.1 |
Pricing clarity reflects the prices the pages leave out and the places where two pages disagree, five for each of the three, plus Deepgram's promotional streaming rates. Deepgram's lower overall score comes mostly from data use and compliance: requests are used for model improvement unless each one opts out, no delete endpoint or DPA with the terms is stated, and the BAA is through sales. AssemblyAI's lead over Speechmatics comes from its scenario price and its self-serve BAA and SSO; Speechmatics scores higher on data use, with training off by default.
Also in this category
Google Cloud Speech-to-Text scores 7.8: dynamic batch at $0.003 a minute puts our scenario at $90 a month (computed), customer data is not used for training without permission, and the BAA, EU processing and SSO come with a Google Cloud account, while PII redaction is not offered and the free minutes apply to the V1 API only. Amazon Transcribe scores 7.0: $0.006 a minute for batch with diarization, custom vocabulary and language identification included ($180 a month for our scenario, computed) and the BAA accepted in AWS Artifact, while voice inputs are used to improve the service unless your organization sets an opt-out policy.
Who should choose which
- Choose AssemblyAI if you need a HIPAA BAA, a training opt-out and retention controls without a sales contract, at the lowest 500-hour price of the three. Look elsewhere if you want training off on the free tier or self-hosted batch transcription.
- Choose Speechmatics if you want customer audio kept out of training by default, short stated retention and a choice of EU, US or Australian processing. Look elsewhere if you need PII redaction or SSO without an Enterprise contract.
- Choose Deepgram if you transcribe large volumes of pre-recorded audio and want speaker labels in the base rate, a $200 credit to test with, and a path to self-hosting. Look elsewhere if you need a BAA or a DPA without an Enterprise contract.
Bottom line
From public pages alone, AssemblyAI edges out Speechmatics, 8.4 to 8.1: AssemblyAI has the lower 500-hour price and compliance controls you switch on yourself, while Speechmatics keeps your audio out of training by default and deletes batch files after 7 days. Deepgram has the largest free credit and speaker labels in its batch rate, with model improvement on unless each request opts out. For all three, price the add-ons and streaming you need, not only the batch rate.
How we checked
We read the vendors' pricing pages and the machine-readable pricing files they link, model, language, billing, rate-limit, data-usage, retention and region docs, trust and security pages, and legal terms on Oct 16, 2026. Speechmatics' Trust Center returned an error from our location, so we used its Internet Archive copy of Dec 3, 2025. We did not sign up, start a trial or call an API, and we did not test accuracy or speed. Facts the pages did not state, such as Deepgram's price for Nova-3 Medical or Speechmatics' discount rate above 500 hours, are left out rather than guessed. Every figure is on the vendors' pages as of that date or, where marked, computed from them.
