Teams that go looking for Deepgram alternatives rarely arrive because the transcripts were bad. They arrive because the transcripts were 97% fine and the 3% that broke were the medication names.

That is a different problem, and it does not get fixed by picking whichever vendor publishes the better accuracy number. Every speech API in this category publishes one. None of them were measured the same way.

So the useful version of this article is not a ranked list. It is the set of checks that decide whether switching is worth doing at all.

Key Takeaway: Deepgram's Nova-3 Medical is a capable clinical speech model, so the reason teams move is rarely raw transcription quality. It is that every vendor publishes a medical accuracy number measured on its own private dataset against an unnamed competitor, that the headline per-minute rate excludes diarization and keyterm prompting, and that the transcript is where these APIs stop. Before switching, benchmark candidates on your own recordings, count errors on drug names rather than on all words, check the data-retention default rather than the compliance badge, and ask how much of the clinical pipeline you are still left to build after the transcript comes back. Clinical APIs built for this workload, Sully among them, return the finished note in your own template and schema-validated fields for the EHR rather than a transcript you still have to process.

Why Teams Running Clinical Audio Start Looking at Deepgram Alternatives

Drug Names Are the Usual Trigger

The failure that starts the search is almost never a bad transcript. It is a good transcript with the wrong drug name in it.

There is a reason for that. A general speech model is tuned for average word accuracy, and drug names are rare tokens. A model can post an excellent aggregate score while missing most of the words that carry clinical risk, because those words are a tiny share of the total.

Deepgram knows this, which is why keyterm prompting is sold as its own capability rather than running by default. It is also capped at 100 custom terms [1][2]. For a product covering several specialties, 100 is a budget you have to spend carefully.

Deepgram Does Have a Medical Model, and It Is Not the Weak Point

Worth saying plainly before anything else: Deepgram's clinical speech model is good.

Nova-3 Medical launched in March 2025, trained on medical terminology covering symptoms, diagnoses, medications and clinical jargon [1]. It runs batch and streaming, accepts keyterm prompting for up to 100 custom terms, and deploys managed, self-hosted or inside your own VPC [1][3]. Deepgram claims it outperforms alternatives by 63.7% on average [3].

If the reason you are leaving is that Deepgram cannot handle clinical audio, check that reason again. It is usually not what is actually wrong. What usually is wrong sits further down: what that accuracy number was measured against, what the bill looks like with clinical features switched on, and where the API stops.

The Cost Shows Up After the Headline Rate

The second trigger is the bill. Nova-3 monolingual streaming starts at $0.0048 per minute on pay as you go [4], which reads cheap right up until you switch on the features a clinical recording needs.

Diarization and keyterm prompting are separate line items [4]. The arithmetic is further down.

Every Vendor Publishes a Medical Accuracy Number and None of Them Compare

Here is the part nobody in this category says out loud. Every speech API on this page publishes a medical accuracy number, and not one of them can be set against another.

Not because anyone is lying. Because each number was produced by the vendor it flatters, on a test set nobody else can see.

What Deepgram Publishes

Deepgram reports a 3.44% median word error rate for Nova-3 Medical, a 6.79% keyword error rate and a 93.99% keyword recall rate [1].

Read the comparison sets, not the figures. The word error rate is framed as a 63.7% improvement over the next-best competitor, and that competitor is never named [1]. The keyword recall figure is measured against Nova-2 Medical, Deepgram's own earlier model, so that one at least has a baseline you can identify [1].

The data behind it is described as an extensive and diverse medical audio dataset drawn from public datasets and proprietary customer audio. No size, no composition, no split, and no stated limitations [1].

The general Nova-3 launch follows the same shape: a 5.26% median batch word error rate against an unnamed next-best competitor at 10% [2].

What Deepgram Competitors Publish

Speechmatics publishes a 1.07% English word error rate against Deepgram's 1.62%, plus 96% medical term recall, SNOMED CT and drug-name training, and up to 50% fewer critical errors [5]. Those figures were measured and published by Speechmatics.

Gladia points readers at Speechmatics for healthcare and publishes no medical accuracy figure of its own [6].

AssemblyAI gives medical transcription and ambient scribes their own section and publishes no medical accuracy number either [7].

There is a pattern in that. The vendors with a medical model publish a medical number. The vendors without one change the subject.

Why You Cannot Stack These Numbers Against Each Other

Four reasons, and all four apply to every figure above.

The test sets are private and different. No vendor here publishes the audio it scored against, so no two numbers describe the same task.

The comparison sets are unnamed. The next-best competitor might be the vendor you are evaluating. It might be one you ruled out a year ago. You cannot tell.

Word error rate weights every word equally. "The" and "hydrochlorothiazide" count the same. A model can win on word error rate while losing on every term that would change a prescription.

Nobody publishes the error breakdown by term type. Medications, dosages, anatomy, clinician names. That breakdown is the only one a clinical product cares about, and it is in none of these announcements.

Sully does not publish a transcription word error rate at all. So this is not a setup for our number to win. It is the reason the only medical accuracy figure that can decide your switch is the one you generate from your own recordings.

What the Deepgram Bill Looks Like Once Clinical Features Are On

The Base Rate for Nova-3

Nova-3 monolingual runs $0.0048 per minute streaming and $0.0043 per minute pre-recorded on pay as you go. On the Growth plan those drop to $0.0042 and $0.0036. Multilingual costs more, at $0.0058 streaming and $0.0052 pre-recorded [4].

Growth starts at $4,000 a year prepaid for up to 20% off, and new accounts get $200 in free credit [4].

Diarization and Keyterm Prompting Are Priced Separately

The two features a clinical recording needs most are the two sitting outside the base rate.

Speaker diarization adds $0.0020 per minute on streaming, though it is included on pre-recorded audio. Keyterm prompting, the feature that fixes drug names, adds $0.0013 per minute on both [4].

Add them up for a streaming clinical encounter with both switched on. $0.0048 plus $0.0020 plus $0.0013 comes to $0.0081 per minute, roughly 69% above the advertised rate [4].

None of that is hidden. It is just not the number you budgeted with.

The Business Associate Agreement Is a Separate Conversation

Deepgram's compliance documentation states that the company is considered a Business Associate under US HIPAA legislation, and that it can provide its Business Associate Agreement to qualifying customers on request, routed through a contact form. It names no plan and no tier [8].

Two things follow. First, "upon request" is a procurement step running on somebody else's timeline, not a checkbox at signup, so put it at the front of your evaluation rather than the end.

The second is easier to miss. The page scopes the offer to customers who qualify as a Covered Entity. A health-tech company building a product is usually a business associate rather than a covered entity, so confirm how the agreement is written for your entity type before you plan around it [8].

The Question That Matters More Than Word Error Rate

A Transcript Is Not a Clinical Note

Your product does not ship a transcript to anyone. It ships a note into a chart, structured fields into an EHR, or a coded claim.

The transcript is an intermediate artifact, and turning one into a usable clinical note is its own problem. Which makes the honest unit of comparison between speech vendors something other than accuracy per word. It is how far past the transcript the API carries you, and how much of the remaining distance your team builds and then maintains.

Where Each Vendor Stops

Deepgram's medical page lists summaries, speaker sentiment, topics and intent. It does not offer clinical note generation, structured output against a schema, or medical coding [3]. One customer testimonial on that page mentions an AI Scribe, but that is the customer's own product built on top of Deepgram's transcription, not a Deepgram feature [3].

Speechmatics, Gladia and AssemblyAI are speech companies too. They end at the transcript and its metadata [5][6][7]. The same holds for Amazon Transcribe Medical, which returns a transcript and points you at other AWS services for everything after it.

That is not a flaw in any of them. It is the category. Speech vendors sell speech.

What You Are Left Building Yourself

Name the work, because you have either done it already or you are about to scope it.

A template engine that turns a transcript into a note in your format. A prompt and evaluation layer on top of a general LLM to handle the clinical reasoning. A schema validator so the output can be written to an EHR without a human reading it first.

Then the parts that are not code. A separate BAA with whichever model vendor you added. A separate security review. A separate uptime risk, and a separate line in your incident runbook.

That second vendor is usually where the clinical risk and the second compliance review actually live. It almost never appears in the speech-API comparison that started the project.

What to Check Before You Switch Transcription Vendors

1. Run Your Own Benchmark on Your Own Recordings

Pull 20 to 50 real encounters covering your actual specialty mix, accents and recording conditions, then run every candidate against the same files.

This is the only comparison on this page that is methodologically sound, and it is cheap. Deepgram gives new accounts $200 in free credit [4] and most competitors have a free tier.

Look at where the traffic goes on this search, too. The most-read result is not a vendor page. It is a Reddit thread from an engineer who benchmarked twelve speech APIs independently, which tells you what this category actually trusts.

2. Count the Drug Names, Not the Words

Score your benchmark on a term list instead of aggregate word error rate. Your top 200 medications, the dosage formats you actually see, your specialty vocabulary, and the clinician and patient names that show up.

A model that posts a better overall word error rate while missing more of that list is the worse model for a clinical product.

This is also the test that tells you whether a keyterm or word-boosting feature is doing anything, which matters when the cap is 100 terms [1][2].

3. Read the Data Retention Default, Not the Compliance Badge

Deepgram's Model Improvement Partnership Program documentation says participation is voluntary, and that opting out means adding mip_opt_out=true as a query parameter to all API requests [9].

Data from opted-out requests is kept only as long as needed to process them. For other requests, Deepgram stores fractional increments of data for model improvement and support, with no retention period stated. The page does not say whether HIPAA or BAA customers are excluded by default [9].

The opt-out exists and it works. The point is that it is per-request and off by default. Find the flag, set it on every request including the ones a teammate adds next quarter, and confirm in writing how it interacts with your BAA.

4. Price the Add-Ons You Will Actually Turn On

Build the quote from the configuration you will ship, not the one on the pricing page. For a streaming clinical encounter with diarization and keyterm prompting, Deepgram's effective rate is $0.0081 per minute against a $0.0048 headline [4].

Do the same for every candidate, then ask what moves at your volume and what a commitment tier costs you in flexibility.

5. Ask How Far Past the Transcript the API Goes

Write down the last artifact your product needs. A formatted note. Validated JSON for an EHR write-back. A coded claim.

Then ask each vendor which of those they return.

The gap between their last artifact and yours is your roadmap, your second vendor, and your second compliance review.

How Sully Handles the Same Workload

You have just been handed five tests. Here is how Sully answers them, including the two where we have no number to offer.

Sully is a research company. It builds and runs its own clinical models, then sells access to them as APIs [10].

Word Boosting, Notes and Structured Fields in One API

Word boosting.Streaming transcription lets you boost the words that matter, such as drug names [10]. That is check two, handled in the transcription call rather than in post-processing.

Recorded audio.Transcription takes recorded visits up to 24 hours long across 30+ languages, batch or streaming, with polling or a webhook when the transcript is ready [10].

Note generation.Send a transcript and your own note template, get the finished note back. Text to JSON takes your JSON Schema and returns validated fields, ready for an EHR [10]. That is check five, answered inside one vendor.

Clinical reasoning. Sully M1 is a clinical LLM called like any chat model, in OpenAI request format, for the work the task APIs do not cover [10]. No second model vendor, no second BAA.

Data handling.Zero Data Retention is the published default, nothing stored after processing, and Sully does not train on customer data. ISO 27001, HIPAA, HITRUST and SOC 2 Type II, with a BAA signed with every customer [10][11]. That is check three, and the contrast with a per-request opt-out flag is the honest version of this point.

Someone has already shipped a product on this. Tebra's AI Scribe is built on Sully, which is the proof point that belongs to the API rather than to the platform.

Across all customers, the APIs handle more than 100,000 notes a month across 40+ specialties at 99.96% uptime and an 11.23 second p50 [10].

What Sully Does Not Publish

On two of the five checks, we cannot hand you a number.

Sully publishes no transcription word error rate, and no per-minute transcription price. So on check one and check four we are asking to be measured rather than offering a figure to copy.

That is the right way round for an article that just spent a section arguing vendor benchmarks are not comparable. Ending it with ours would be strange.

Run the test instead. Get an API key at dashboard.api.sully.ai, point it at 20 of your own recordings, and count the drug names. Change the base URL and the model name. There is no demo to sit through.

FAQ

How expensive is Deepgram for medical transcription?

Deepgram's Nova-3 monolingual model starts at $0.0048 per minute for streaming and $0.0043 per minute for pre-recorded audio on pay as you go, with lower rates on the Growth plan, which begins at $4,000 a year prepaid [4]. The clinical configuration costs more than the headline: speaker diarization adds $0.0020 per minute on streaming and keyterm prompting adds $0.0013 per minute, so a streaming encounter with both enabled runs about $0.0081 per minute [4]. New accounts receive $200 in free credit, which is enough to benchmark your own audio before committing [4].

Does Deepgram have a medical speech to text model?

Yes. Nova-3 Medical launched in March 2025, trained on medical terminology including symptoms, diagnoses, medications and clinical jargon, and called with model=nova-3-medical [1]. It supports batch and streaming audio, keyterm prompting for up to 100 custom terms, and managed or self-hosted deployment including on-premises and VPC configurations [1][3]. Deepgram publishes a 3.44% median word error rate for it, described as a 63.7% improvement over an unnamed next-best competitor [1]. The model is a transcription model, so it returns a transcript and its metadata. If your product needs the finished note or EHR-ready fields, that work sits outside it [3][10].

Is Deepgram HIPAA compliant and will it sign a BAA?

Deepgram's compliance documentation states that it is considered a Business Associate under US HIPAA legislation and can provide its Business Associate Agreement to qualifying customers on request, through its contact form [8]. The page names no plan or tier, and it scopes the offer to customers who qualify as a Covered Entity, so a health-tech company that is itself a business associate should confirm how the agreement is written for its entity type. Separately, check the Model Improvement Partnership Program default: opting out requires adding mip_opt_out=true to all API requests [9].

Which is better for clinical audio, AssemblyAI or Deepgram?

Deepgram publishes a dedicated medical model with medical accuracy figures. AssemblyAI discusses medical transcription and ambient scribes but publishes no medical accuracy number [1][7]. On that evidence Deepgram is the stronger clinical speech model of the two, but both numbers and both absences are vendor-controlled, so settle the comparison on your own recordings rather than on either company's page. Score the test on your medication and specialty term list rather than on overall word error rate.

Can a transcription API write the clinical note, not just the transcript?

Most cannot. Deepgram, Speechmatics, Gladia and AssemblyAI are speech companies whose output ends at the transcript and its metadata, such as summaries, topics, sentiment and speaker labels [3][5][6][7]. Sully's APIs continue past that point: note generation returns a finished note from a transcript and your own template, Text to JSON returns schema-validated fields ready for an EHR, and Sully M1 is a clinical LLM for the reasoning the task APIs do not cover [10]. If your product ships a note rather than a transcript, that gap is the part of the build the speech-API comparison never shows you.

Sources

[1] Deepgram — Introducing Nova-3 Medical: Speech-to-Text for Healthcare [2] Deepgram — Introducing Nova-3: Speech-to-Text API [3] Deepgram — Medical Transcription Speech to Text API [4] Deepgram — Pricing [5] Speechmatics— Deepgram Alternative: Speechmatics vs Deepgram [6] Gladia — 7 Deepgram Alternatives: Speech-to-Text Solutions for Specific Business Needs [7] AssemblyAI— 5 Deepgram Alternatives in 2026 [8] Deepgram — Data Privacy Compliance [9] Deepgram — The Deepgram Model Improvement Partnership Program [10] Sully.ai — Sully API [11] Sully.ai— Trust Center