Best Speech-to-Text APIs, Priced at Real Monthly Volumes

The best speech-to-text application programming interface (API) comes down to one question: does anything need to read the transcript while people are still talking?

At Layer3Labs, we build call and intake automation for clients. That means we judge these tools by the errors that actually make it into a workflow, not by where they sit on a leaderboard.

For live products, Meta's Muse Voice Transcribe is the cheapest published option that includes speaker labels, while Deepgram Nova-3 offers the most control. For finished recordings, Deepgram's pre-recorded rate is hard to beat. And if the audio can't leave your servers, self-hosted Whisper is the only real answer.

Volume can change the ranking. At 10 hours a month, the price differences are small enough to ignore. At 1,000 hours, the gap between the cheapest and most expensive options with speaker labels grows to $390 a month.

All rates are the companies' own published figures at the time of writing, with links to the original pages. Deepgram's streaming rate was promotional as of September 1, 2026, and is marked accordingly.

Deepgram Nova-3 vs. Muse Voice Transcribe: Side-by-Side

How We Ranked These Speech-to-Text APIs

We ranked on four things a transcription workflow depends on, in this order.

Raw accuracy sits deliberately low on the four criteria listed at the end of this section. Every company here publishes a word error rate on a test set of its own choosing, and the spread between the serious options is smaller than the spread between your best and worst recordings. A model that scores two points better on clean benchmark audio can lose badly on a phone line with crosstalk.

We did not rank on feature checklists. Long lists of supported formats and callbacks look decisive on a comparison page and rarely change which vendor a team keeps after three months.

Prices below are per hour of audio unless stated. Where a company prices per minute we converted it and said so. Promotional rates are flagged, because a promotional rate is a discount with an expiry rather than a price.

• Whether the API works in the stream, because a batch-only model rules out every live product.
• Whether speaker labels come with it, since buying them separately can double the hourly rate.
• What a real month costs at your volume, which changes the ranking between 10 hours and 1,000.
• Whether you can run it where your data has to stay, which overrules every other criterion when it applies.

What Each API Costs at 10, 100 and 1,000 Hours a Month

Monthly totals for streaming transcription with speaker labels, at each company's published rate:

At 10 hours a month, every difference in that table is smaller than one lunch. Choose on fit and ignore price entirely. At 1,000 hours, the gap between the cheapest and dearest option with labels is $390 a month. That is worth a migration for a support desk, and smaller than most teams expect.

The self-hosted line is the one that misleads people. Running Whisper on your own hardware has no per-hour bill. A machine capable of real-time transcription, plus the engineering hours to keep it running, usually costs more than the $180 a month that 1,000 hours buys on the cheapest hosted option.

→ Compare per-token pricing for every AI model

Best for Live Transcription and Voice Agents

Muse Voice Transcribe is the strongest live option on published rates, because it does three jobs in one call. The Meta model transcribes, attributes speech to 20 or more speakers while the audio is still arriving, and detects when a speaker has finished. It costs $3.00 per 1,000 audio minutes, or $0.18 per hour.

A voice agent needs that last capability. An agent has to know when the caller stopped talking, and getting that signal from the same model that wrote the words removes a component and a source of timestamp disagreement. Meta reports a final transcript about 0.16 seconds after end of speech, and a 3.1% word error rate on the Artificial Analysis streaming benchmark, as reported on September 1, 2026.

Deepgram Nova-3 is the alternative worth trialling in the same week. It supports diarization on streaming through a parameter, its endpointing settings are tunable per request, and it has years of production traffic behind it. It costs about two and a half times as much at its standard rate.

The tuning difference shows up in production. Different call types want different silence thresholds, and a noisy line with long pauses will either start a new speaker turn or not. Deepgram lets you set that threshold per request. Muse Voice Transcribe decides it for you.

Best for Batch Transcription of Finished Recordings

The Deepgram pre-recorded tier is the value pick for finished audio, listed at $0.0043 per minute for Nova-3, about $0.26 per hour. Speaker diarization is included, and confidence scores come back alongside the labels.

Those confidence scores matter more than they sound. They let you route only the uncertain turns to a person for review, so a reviewer reads the handful of segments the model was unsure about instead of every transcript. Streaming diarization does not return them.

AssemblyAI is the pick when you want more than a transcript from the same call. Its async tiers run $0.21 per hour on Universal-3.5 Pro and $0.15 per hour on its lower tier, with diarization at $0.02 per hour. It also layers summarisation and topic detection on the same audio.

The OpenAI batch models fit teams already building there. gpt-4o-transcribe runs $0.006 per minute, about $0.36 per hour, gpt-4o-mini-transcribe runs $0.003 per minute, and gpt-4o-transcribe-diarize adds speaker labels at $0.006 per minute. One vendor and one invoice saves real administrative time on a small team.

Best for Regulated Audio and Self-Hosting

Whisper is the answer when audio cannot leave your infrastructure, and it is the only real answer on this list. The weights are open, it runs on your own hardware, and no recording crosses a network boundary you do not control.

It has no speaker labelling at all, so teams pair it with an open library such as pyannote and align the two outputs by timestamp. That merge produces most attribution errors, at the turn boundaries where the two models disagree, and speaker diarization covers how to keep them down. It is also engineering you own forever rather than a line on an invoice.

Google Cloud Speech-to-Text is the option when procurement rather than engineering is the constraint. If your organisation already holds a Google Cloud contract with the regions and terms your compliance team approved, buying transcription inside it can clear review faster than a better model from a new vendor. Google publishes per-minute rates that vary by model and volume on that pricing page.

In the client intake workflows we have automated for law firms, the deciding factor was almost never the model. It was whether the recording could sit on a third-party server at all, and a contract answers that question rather than a benchmark.

Who Should Skip All of These

Two groups are better served elsewhere, and they are named at the end of this section.

Three things would change this ranking. Deepgram could match the bundled rate. Meta could ship a self-hosted option or publish weights. Six months of production reports could show the newest model behaving differently under real concurrency than in a trial. Any of the three changes which option leads for live transcription. To pick between the top two on your own recordings rather than on published rates, book a workflow audit.

• Teams who need finished meeting notes rather than a transcript stream. A packaged tool that already records, transcribes, summarises and syncs to a customer relationship management (CRM) system gets you there faster, because these APIs are components rather than products.
• Teams transcribing a few hours a month. The free tiers built into tools you already pay for will cover it, and writing a client around an API returns nothing at that volume.

The Verdict

For live transcription with speaker labels, Muse Voice Transcribe is the cheapest published option and the simplest stack, at $0.18 per hour with attribution and end-of-speech detection in the same model. Its weakness is its age. Meta announced it on September 1, 2026, so it has no production history, and there is no self-hosted path.

For production streaming that needs per-request control and a longer track record, Deepgram Nova-3 is worth its higher rate. For finished recordings, the Deepgram pre-recorded tier at about $0.26 per hour with diarization and confidence scores is the value pick. For audio that cannot leave your servers, self-hosted Whisper with pyannote is the only real answer.

Below about 100 hours a month, ignore the price table and pick on fit. Above 1,000 hours, run your own worst hour of audio through your top two and count the errors that would have reached your team, because that number decides more than any published benchmark.