Author(s): Mia Efoxtech

Originally published on Towards AI.

The Cheapest Speech-to-Text API Can Be the Most Expensive

The Cheapest Speech-to-Text API Can Be the Most Expensive

The Cheapest Speech-to-Text API Can Be the Most Expensive

In 2026, production transcription cost is shaped by channels, feature fees, latency, retries, and human correction — not the advertised price per minute.

The lowest speech-to-text list price does not guarantee the lowest production cost. A $0.15-per-hour API can cost more than a $0.60 route after channel billing, retries, paid features, missed deadlines, and editor correction time are included. The useful metric is cost per accepted transcript, measured on your own audio.

Key takeaways

• AssemblyAI Universal-2 starts at $0.15 per recorded audio hour, while Google Dynamic Batch and OpenAI gpt-4o-mini-transcribe normalize to about $0.18 per hour.
• Those prices are not equivalent: Dynamic Batch is lower priority, channel billing differs, and feature fees vary by provider.
• Real-time products should be measured on stable partials and finalization delay, not only time to first text.
• Human correction can exceed the API charge by an order of magnitude.
• The right production design may use separate routes for live calls, multilingual meetings, and non-urgent archives.

A price table answers the wrong question

Speech-to-text pricing looks unusually easy to compare. Most providers publish a price per minute or hour, and the numbers can be normalized into one clean table.

That table is useful for building a shortlist. It is a poor basis for signing a production contract.

As of July 27, 2026, the lowest public recorded-audio prices in this comparison are clustered between $0.15 and $0.22 per audio hour. AssemblyAI Universal-2 is listed at $0.15 per hour. Google Dynamic Batch and OpenAI gpt-4o-mini-transcribe normalize to $0.18 per hour. ElevenLabs Scribe v2 is $0.22 per hour.

The numbers look close. The products are not.

Google Dynamic Batch trades turnaround priority for price. AssemblyAI bills multichannel files per channel. OpenAI uploaded transcription has a 25 MB file limit. ElevenLabs charges separately for entity detection and keyterm prompting. A one-line price comparison hides the conditions that determine whether the route fits the workload.

The shortlist is not a ranking. It is a way to decide which providers deserve the same test audio.

Channel billing can reverse the result

A 60-minute stereo call is not always billed as one audio hour.

AssemblyAI and Google bill each audio channel separately in the routes described here. AWS includes up to two channels in the duration charge for standard transcription. That difference matters for contact-center audio, where one speaker may be isolated on each channel.

Using the July 27 public prices and the AWS provider example cited in the source research:

The AWS value is an official high-volume US East example at two million monthly minutes, not a universal entry quote. The table also says nothing about overlapping speech, speaker attribution, redaction, or correction effort.

That is precisely the point. Billing rules can change the ordering before transcript quality is measured.

Add-ons belong in the base case

Speaker diarization, redaction, keyterm prompting, and entity detection are often presented as optional features. For many products, they are requirements.

Deepgram’s Nova-3 Monolingual pre-recorded route rises from $0.0077 to $0.0097 per minute when diarization is added. Adding redaction raises the same route to $0.0117 per minute before keyterm prompting. AssemblyAI lists recorded diarization at $0.02 per hour and real-time diarization at $0.12 per hour. ElevenLabs lists entity detection at $0.07 per hour and keyterm prompting at $0.05 per hour.

The honest planning formula is:

Cost per accepted transcript =
API processing
+ channel charges
+ paid features
+ retries
+ storage and transfer
+ human correction time

A procurement spreadsheet should model required features by workload rather than placing them in a footnote.

Human correction is usually the largest hidden line

Transcription is rarely valuable because a provider returned a JSON response. It becomes valuable when a person, workflow, or downstream model can trust the result.

Consider a deliberately simple planning example. Assume the reviewer costs $30 per hour:

The correction times are assumptions, not provider benchmark results. They show why teams must replace list-price comparisons with their own review measurements.

In this example, the second API is four times more expensive at the processing layer and roughly half the cost after review. A small improvement in names, punctuation, speaker labels, or number formatting can be worth much more than the difference between two API rates.

Real-time transcription needs a different scorecard

Recorded transcription and live transcription should not share one benchmark report.

A batch API can be evaluated on upload-to-final turnaround, formatting, long-file behavior, and accepted-transcript cost. A voice agent needs stable partial results, endpointing, interruption handling, reconnect behavior, and finalization delay.

Time to first text is not enough. A system can display an early partial quickly and then make the user wait for a stable final result. Measure first partial, stable partial, and final transcript at both p50 and p95.

The same applies to latency claims. Vendor model latency excludes the microphone path, region, network, WebSocket client, buffering, UI update, and downstream agent. Production latency is the complete path that a user experiences.

Accuracy should be tested on the audio you actually own

Public accuracy claims are difficult to compare because providers use different datasets, language mixes, normalization policies, acoustic conditions, and model versions.

The 2026 GigaSpeechBench study evaluates 680 hours of real-world speech across low-resource languages, Chinese dialects, English accents, 12 vertical domains, and child and older-adult speech. The study reports substantial degradation for leading models and commercial APIs in difficult settings.

The durable conclusion is not that one public benchmark identified a permanent winner. The conclusion is that a useful test set must resemble production traffic.

A practical evaluation set should contain:

• clean meetings with names and domain terminology
• noisy support calls with crosstalk and hold music
• accented and multilingual speech with genuine code switching
• long-form podcasts or video files
• short live utterances with hesitation, silence, and barge-in

Word error rate belongs in the report, but it should sit beside named-term accuracy, speaker attribution, formatting, p95 latency, reliability, and editor minutes per audio hour.

A seven-day evaluation is better than a permanent ranking

The following process is usually enough to replace opinion with evidence:

1. Define required languages, timestamps, p95 latency, speaker-label tolerance, data terms, and maximum review time.
2. Build a licensed, production-like audio set with one reference-transcript policy.
3. Normalize regions, audio formats, channel layouts, vocabulary hints, retries, and model versions.
4. Run recorded-audio tests and measure cost, turnaround, accuracy, failures, and correction time.
5. Run separate live tests for stable partials, finalization, endpointing, interruptions, and reconnects.
6. Calculate cost per accepted transcript with channels, add-ons, retries, storage, and reviewer time.
7. Choose a routing policy instead of forcing one provider to handle every workload.

Teams evaluating supported OpenAI-compatible speech models can use a unified route such as the CometAPI Quickstart to keep authentication and client setup consistent. The exact model ID, audio limits, feature support, and privacy terms still need to be verified for every route.

The right answer is usually a routing policy

There is no universal best speech-to-text API in 2026.

AssemblyAI and Google Dynamic Batch are sensible low-cost candidates for recorded archives. Deepgram, AssemblyAI, and ElevenLabs belong on early real-time shortlists. OpenAI is convenient when the surrounding application already uses OpenAI-style tooling. AWS and Google can reduce operational friction when the data, identity, logging, and procurement stack already live in their clouds.

The most defensible production design may use one provider for live English calls, another for multilingual meetings, and a lower-priority batch route for archives.

The goal is not to pick the winner of a pricing table. The goal is to produce accepted transcripts inside the product’s cost, latency, quality, and data constraints.

FAQ

What is the cheapest speech-to-text API for recorded audio?

AssemblyAI Universal-2 starts at $0.15 per audio hour in the July 27, 2026 public-price snapshot. Google Dynamic Batch and OpenAI gpt-4o-mini-transcribe normalize to about $0.18 per hour. Channel charges, add-ons, retries, storage, and human review can change the final ordering.

Which speech-to-text API is best for real-time transcription?

Start by testing Deepgram, AssemblyAI, and ElevenLabs. Include OpenAI, AWS, or Google when ecosystem fit or language support matters. Measure stable partials, finalization delay, endpointing, interruption handling, reconnects, and p95 latency on the actual transport path.

Is word error rate enough for choosing a provider?

No. Word error rate does not measure speaker attribution, proper nouns, formatting, finalization delay, reliability, or correction time. Use one normalization policy and report WER alongside named-term accuracy, speaker errors, p95 latency, failures, and editor minutes per audio hour.

Should a team use one speech-to-text provider for every workload?

Not necessarily. Recorded archives, live voice agents, multilingual meetings, and contact-center calls have different requirements and billing rules. A tested routing policy can be more reliable and economical than forcing one route to handle every audio type.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI