If you have read two or three guides ranking the best speech to text API, you may have noticed something odd. Each one puts a different API at number one, and in almost every case it is the API sold by the company that wrote the guide.

That is not a conspiracy. It is just what happens when the only people motivated to write a 4,000-word comparison are the vendors being compared.

It leaves the engineer doing the evaluation with six rankings that disagree, no shared test audio, and no way to check any of it.

This article is published by a vendor too. So it does three things differently: it names the conflict up front, it states the scoring criteria before the ranking instead of after, and it ends by handing you the test to run on your own recordings.

Key Takeaway: Most published rankings of the best speech to text API are written by speech to text vendors, and in five of the six guides ranking for the term, the publisher places its own API at or near the top. For clinical audio the comparison turns on four things the general rankings do not score: whether you can boost your own drug and specialty terms in the API call, whether the transcript, the note and the structured fields come from one vendor, what the default data retention posture is, and how far past the raw transcript the API goes. Sully's APIs answer all four on one key, with word boosting on streaming transcription, recorded visits up to 24 hours across 30+ languages, note generation from your own template, schema-validated fields for the EHR, and Zero Data Retention by default.

Why Almost Every Speech to Text API Ranking Is Published by a Vendor

Six comparison guides rank on the first page of Google for this term. Every one of them is published by a company selling in the category.

In five of the six, the publisher's own product sits at or near the top.

Who Publishes the Guides That Rank for This Term

Gladia's guide ranks Gladia first, and sources the leadership claim to Gladia's own benchmarks [1]. Speechmatics puts Speechmatics first, with more sub-detail on its own entry than on any other vendor's [2]. Inworld ranks its Realtime STT first and the page calls it "the strongest speech-to-text API in this comparison" [3].

Voicy's guide lists ten APIs, then closes by recommending Voicy instead of all of them [4]. AssemblyAI's entry is a free-tier guide built around AssemblyAI [5]. The sixth, Eden AI, ranks nobody at all and names no provider other than Eden AI [6].

Not one of the six scores anyone on clinical audio. Medical shows up as a passing feature mention in three of them and nowhere in the other three.

The Result That Outranks All of Them Is an Independent Benchmark

The highest-ranking organic result for this term is not a guide. It is a Reddit thread by an engineer who benchmarked twelve speech to text APIs himself.

It pulls roughly 326 organic visits a month. The best-performing vendor guide on the page pulls 76. It is also the lead citation in Google's AI Overview for the term, and the AI Overview cites no vendor guide at all.

Read that as a verdict. Given the choice between six professionally written comparisons and one engineer posting his own numbers, both readers and Google's own summary pick the engineer.

What This Article Does Differently

Sully sells these APIs. That is the bias, stated plainly.

What a vendor can still do is make its ranking checkable. The next section lists the four criteria before any vendor is named, so you can argue with the criteria rather than guess at them. The section after the ranking hands you the test to run on your own audio, which is the only ranking that counts.

The Four Things That Decide Speech to Text Accuracy on Clinical Audio

General speech to text accuracy is close enough between the leading APIs that it rarely decides a product. On clinical audio, four narrower things do.

Whether You Can Boost Your Own Drug and Specialty Terms

A model trained on general speech has no reason to prefer "metoprolol" over a common English word that sounds like it. The fix is a list of your own terms passed into the API call, not a find-and-replace afterwards.

Every serious vendor offers some version of this, and every one of them calls it something different. Deepgram calls it keyterm prompting and charges $0.0013 a minute for it [7]. AssemblyAI calls it keyterms prompting, included on some streaming models and $0.05 an hour on async [8]. Speechmatics calls it a custom dictionary, available on two of its models [9]. Sully calls it word boosting, on streaming transcription[10].

The naming alone is part of why these guides are hard to compare. Four vendors, four words for one feature, and three different billing treatments.

Whether the Whole Pipeline Comes From One Vendor

A clinical note has to know who said what. The clinician's words and the patient's words land in different parts of the note, and a transcript that cannot tell them apart is not usable without a second pass.

So check two things together. Does the API do speaker diarization, and does it do it in a live stream?

They are not always the same answer. On OpenAI's transcription path, diarization runs only through one model and is not supported in realtime sessions at all [11]. Deepgram includes diarization free on pre-recorded audio and charges $0.0020 a minute for it on streaming [7]. AssemblyAI charges $0.02 an hour on async and $0.12 an hour on streaming, a six-fold difference for the same feature [8].

When diarization is missing from the streaming path, teams usually bolt on a separate library. That puts another dependency inside the boundary your compliance review has to cover.

What the Default Retention Posture Is

"Will you sign a BAA" is the question everyone asks, and nearly every major vendor says yes. It is close to useless as a differentiator.

The better question is what happens to the audio by default, before anyone negotiates anything.

Sully's APIs run Zero Data Retention, with nothing stored after processing and no training on customer data [10], under ISO 27001, HIPAA, HITRUST and SOC 2 Type II [12]. For the other vendors here, read their own data pages on the day you evaluate, because these defaults change and a guide written six months ago is not evidence.

How Far Past the Transcript the API Goes

Every API on this page returns a transcript. None of them returns a clinical note.

What sits between those two things is a speaker-attributed transcript, terminology checked against your own vocabulary, structured fields validated against your schema, and finally a write into the EHR. Each stage you do not get back from the API is code someone on your team writes and then maintains.

That gap is the real cost comparison, and it never appears on a pricing page.

The Best Speech to Text APIs for Medical Audio

Ranked on the four criteria above. The bias is disclosed, the criteria are published, and prices were pulled from each vendor's own pricing page on 2 November 2026.

1. Sully

Five APIs on one key, built for clinical audio rather than adapted to it. Transcription takes a recorded visit up to 24 hours long across 30+ languages. Streaming transcription runs during the encounter and lets you boost the words that matter, such as drug names. Note generation takes a transcript and your own template and returns the finished note. Text to JSON takes a note with your JSON Schema and returns validated fields. Sully M1 is a general clinical model, called like any chat model, for what the task APIs do not cover [10].

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].

Sully does not publish a per-minute rate, so it is absent from the price chart below. Better said before the chart than after it.

Tebra's AI Scribe is built on these APIs, which is the part worth checking: another health-tech company already shipped a product on them [10].

2. Deepgram Nova-3 Medical

Keyterm prompting up to 100 custom terms, batch and streaming, and deployment managed, self-hosted or inside your own VPC [7]. Nova-3 Medical is trained on medical terminology and covers both batch and streaming, which not every medical model here does.

Nova-3 runs $0.0043 a minute batch and $0.0077 streaming [7]. Nova-3 Medical does not appear on the pricing page at all, so the medical rate is a sales conversation.

Worth reading the accuracy claims carefully. They are Deepgram measuring Deepgram against a competitor it does not name.

3. AssemblyAI

The cheapest general rate on this list. Universal-2 runs $0.15 an hour async and Universal-Streaming English $0.15 an hour, which works out to about a quarter of a cent a minute [8].

Medical Mode is an add-on rather than a separate model, switched on with "domain": "medical-v1"for another $0.15 an hour, and it works on both async and streaming models [8]. Keyterms prompting handles the rest of your vocabulary, and the add-on pricing for diarization is where the streaming bill grows.

4. Speechmatics

The one to look at if you cannot send audio off your own infrastructure. On-premises and on-device deployment, speaker diarization, and a custom dictionary [2][9].

Standard runs $0.45 an hour and the Oak 1 medical model $0.30 an hour, about $0.005 a minute [9]. Speechmatics is the only vendor here whose medical model is cheaper than its general one. Diarization and custom dictionary carry no separate charge, which is unusual on this list, but they are only available on some models [9].

5. Google Cloud Speech-to-Text

The default answer, and the page that owns this search result at domain rating 100. Broad language coverage and mature tooling.

Standard recognition is $0.016 a minute, with dynamic batch at $0.003 for workloads that can wait [13]. Google does have medical models, medical_conversation and medical_dictation, and they cost $0.078 a minute [13]. That is 4.9 times the standard rate, and no guide on this search result mentions it.

6. Amazon Transcribe Medical

Purpose-built for medical speech and the narrowest in scope of anything here. US English only, six specialties, and only primary care accepts an uploaded file.

Standard Transcribe is $0.006 a minute batch and $0.01 streaming [14]. Transcribe Medical works out to $0.075 a minute from Amazon's own worked examples, which price a 60 minute consultation at $4.50 [14]. That is 12.5 times the standard batch rate.

7. OpenAI gpt-transcribe

The replacement path now that whisper-1 is being removed. gpt-transcribe is $0.0045 a minute and gpt-live-transcribe $0.017 [15].

No clinical variant, and diarization is available only through a model on the removal list [11]. Cheap, capable, and the furthest of anything here from a clinical note.

8. ElevenLabs Scribe v2

Scribe v2 Medical is fine-tuned for clinical audio and ElevenLabs reports 35% fewer transcription errors on it than on general Scribe v2, at the same price as the general model [16]. Keyterm prompting is a paid add-on at roughly $0.05 an hour [1].

The constraint to know: Scribe v2 Medical is a batch model. Live transcription runs on the general Scribe v2 Realtime, so an ambient clinical product does not get the medical model during the visit [16].

What Speech to Text API Pricing Looks Like at Clinical Volume

The per-minute rates above are close enough that price should not decide this. Then you look at what each vendor charges for the medical version.

Seven of the eight APIs here now have a medical model. What separates them is the markup.

Google charges $0.078 a minute for its medical models against $0.016 standard, which is 4.9 times [13]. Amazon works out to $0.075 against $0.006 batch, which is 12.5 times [14]. Both are general cloud vendors charging a clinical premium.

The speech specialists do not. Speechmatics prices its Oak 1 medical model at $0.30 an hour against $0.45 for standard, so the clinical version is the cheaper one [9]. AssemblyAI adds $0.15 an hour to a $0.15 base, doubling off a low number [8]. ElevenLabs charges the same for Scribe v2 Medical as for general Scribe v2 [16]. Deepgram has Nova-3 Medical and publishes no rate for it at all [7].

Having a medical model stopped being the differentiator some time in the last year. What it costs, whether it runs during a live visit, and what comes back past the transcript still are.

The Published Rate and the Configuration You Will Actually Ship

Build the quote from what ships, not from the headline.

A clinical encounter is usually streaming, with diarization on and a term list loaded. On Deepgram that is $0.0077 for streaming, plus $0.0020 for diarization, plus $0.0013 for keyterm prompting, which is $0.0110 a minute against a $0.0043 headline [7]. The configuration costs 2.6 times the number you first saw.

On AssemblyAI, streaming diarization is $0.12 an hour on top of a $0.15 base, nearly doubling it [8].

One more line worth reading twice: AssemblyAI bills streaming by WebSocket session time, not by audio sent [8]. An open connection during a quiet exam room still bills.

Where the Per-Minute Comparison Breaks Down

Two of the guides ranking for this term publish different per-minute prices for the same Deepgram model. One says $0.0043 [1], the other $0.0077 [2].

Both are right. The first is the batch rate and the second is the streaming rate, and neither guide says which [7]. An engineer reading both comes away thinking one of them is wrong.

Half these vendors price per hour and half per minute, the medical tiers sit on separate pages, and the add-ons that clinical use requires are billed separately. Read the vendor's pricing page on the day you budget, and write down which mode you are quoting.

How to Check Any of This on Your Own Audio

Nothing above substitutes for testing on your own recordings. Three steps.

1. Build a Term List From Your Own Prescribing Data

Pull the 200 drugs and terms your product actually sees, not a generic clinical word list. The errors that matter are the ones specific to your specialty mix.

2. Score the Terms, Not the Whole Transcript

An aggregate word error rate hides the failures that carry clinical risk, because one missed drug name in a long paragraph barely moves the number.

3. Test the Configuration You Will Ship

Run streaming with diarization and your term list switched on. The batch demo endpoint is not the thing you are buying.

How Sully Handles the Same Workload

Scored on the same four criteria, in the same order.

Term boosting. Streaming transcription lets you boost the words that matter, such as drug names, inside the transcription call rather than in post-processing [10].

One vendor for the pipeline. Streaming transcription returns partial and final transcripts live during the visit, and recorded visits up to 24 hours process batch or streaming with polling or a webhook [10]. The transcript, the note and the structured fields come back from the same API key, so there is no second vendor inside the boundary your compliance review has to cover.

Data handling. Zero Data Retention is the published default, nothing stored after processing, and no training on customer data [10], under ISO 27001, HIPAA, HITRUST and SOC 2 Type II [12].

Past the transcript. Note generation takes a transcript and your own template and returns the finished note. Text to JSON takes your JSON Schema and returns validated fields, ready for an EHR [10]. That is the part of the gap you would otherwise build.

Run the test instead of taking the ranking. Get an API key at dashboard.api.sully.ai, point it at 20 encounters from your own product, and check the medication names against the audio. There is no demo to sit through [10].

FAQ

Which AI has the best speech-to-text?

There is no single answer, because the published rankings disagree and almost all of them are written by one of the vendors being ranked. On general English audio the leading APIs are close enough that the difference rarely decides a product. On clinical audio the deciding factors are narrower: whether you can pass your own drug and specialty terms into the API call, whether diarization and streaming come from the same vendor, and what the service retains by default. Sully's APIs were built for that case, with word boosting on streaming transcription and note generation from your own template [10].

Is there a free API for converting audio to text?

Yes, with limits. Google Cloud Speech-to-Text gives 60 free minutes a month on the V1 API [13], and Amazon Transcribe gives 60 minutes a month free for the first 12 months [14]. The open-source Whisper weights are free to download and run on your own hardware, which removes the per-minute charge and replaces it with GPU, storage and DevOps costs that only pay off at high volume. None of the free tiers are sized for a production clinical workload, and free tiers are usually where the data handling terms differ most from the paid ones, so read those before you test with real audio.

What is the best speech to text API for medical audio?

It depends on what you need past the transcript. For a drop-in transcript the general APIs are close substitutes for each other. For a clinical product the checks that decide it are term boosting, diarization and streaming from one vendor, the default retention posture, and how much of the work between a transcript and an EHR-ready field the API does for you. Sully covers all four on one key, which is why it leads this list, and this article states that bias rather than hiding it [10].

Do speech to text APIs handle drug names correctly?

Not by default. A general model has no reason to prefer a rare drug name over a common English word that sounds like it, and an aggregate word error rate will not surface the problem because one missed word barely moves the score. Every serious API offers some way to pass in your own terms, though each vendor calls it something different: keyterm prompting [7], keyterms prompting [8], custom dictionary [9], word boosting [10]. Measure on your own term list rather than on the vendor's benchmark.

Is a speech to text API HIPAA compliant?

Compliance is a property of the agreement and the configuration, not of the API itself. Most major vendors will enter into a Business Associate Agreement for qualifying customers, so the more useful question is what the service retains by default before anyone negotiates. Sully's APIs run Zero Data Retention, with nothing stored after processing and no training on customer data [10], under ISO 27001, HIPAA, HITRUST and SOC 2 Type II [12]. Ask every vendor the same question and get the answer in writing.

Sources

[1] Gladia — Best Speech-to-Text APIs in 2026 [2] Speechmatics— Best speech-to-text AI guide: APIs, platforms and services compared [3] Inworld — Best Speech-to-Text APIs for Voice AI Developers 2026 [4] Voicy— Best Speech to Text APIs for Developers in 2026 [5] AssemblyAI — Best Free Speech-to-Text APIs and Open-Source Engines[6] Eden AI — Best Speech-to-Text APIs in 2026: Features, Pricing, and Best Use Cases [7] Deepgram — Pricing [8] AssemblyAI — Pricing [9] Speechmatics— Pricing [10] Sully.ai — Sully APIs [11] OpenAI — File transcription [12] Sully.ai — Trust Center [13] Google Cloud — Speech-to-Text pricing [14] Amazon Web Services — Amazon Transcribe pricing [15] OpenAI — API pricing [16] ElevenLabs — Speech to Text