Streaming Speech-to-Text APIs for Devices

Concept render

As of October 31, 2026, AssemblyAI bills streaming by open WebSocket time, so a pendant open 10 hours costs $1.50 a day at $0.15/hr. If only 2 hours of speech are billed, AWS is $1.20 and Deepgram Nova-3 about $0.92. Neither vendor page confirms that idle seconds are free.

Picking the transcription service behind a recorder or pendant depends on what the provider counts: open socket time, audio seconds, new streams per minute or concurrent connections. A device that holds a connection open all day behaves differently under each of those counts. This guide puts the published limits and rates side by side, works through the arithmetic with labeled assumptions, and lists what to measure on your own board. It contains no Duobay price and no measured result.

This page covers consented recorders and pendants only. A recording device should show a visible indicator while it records and should have a clear way to obtain consent. Facts below come from each vendor's own public pages, read on October 31, 2026. Prices and quotas change often, so re-read the linked pages before you decide.

How do the five providers differ on billing unit and limits?

This table keeps only what the vendor pages state. Where a page did not give a figure, the cell says so, and the figure is yours to test.

AWS also publishes the quota rows on its service quotas page and the hard session limit on its streaming documentation page, both as of October 31, 2026. Language support lists differ by provider and change often. We did not capture them, so read each vendor page at design time.

What does the same day of use cost under each billing unit?

The example is a pendant whose socket stays open 10 hours a day while the wearer produces 2 hours of speech. Both hour counts are assumptions for you to replace. The rates come from the table above.

open-time billing   daily cost = H_open   × r_hr
audio-time billing  daily cost = H_speech × 60 × r_min

H_open   = 10 hours   (assumption)
H_speech = 2 hours   (assumption)

Check on the arithmetic: $0.15 × 10 = $1.50 and $0.15 × 2 = $0.30. $0.01 × 120 minutes = $1.20. $0.0077 × 120 = $0.924, shown as $0.92. The audio-time rows assume the provider bills only the 2 hours of speech. For Amazon Transcribe, the docs say to send silence as zero bytes to keep the stream real-time, and the pages we read do not say whether those seconds are billed. For Deepgram, the pricing page does not state streaming granularity separately. Confirm both with the vendor, and measure them against your own invoice.

The lower AssemblyAI figure shows the effect of closing the socket in silence, for example with a voice-activity detector on the device. How often the detector fires, and how much speech it clips, is a design question for your own tests.

How many devices fit under each concurrency limit?

A concurrency limit covers only the devices streaming at the same moment.

concurrent demand = F × a
largest fleet that fits under the limit = limit ÷ a

a = 0.1 (share of devices streaming in the same minute, placeholder)

Two providers count differently. AssemblyAI limits new streams per minute (100 on pay-as-you-go), so the question is how many devices open a stream in the same minute. Google limits recognition requests per project (900 per 60 seconds), and each streaming session is one request.

For Google, the reconnect arithmetic is: sessions per hour of continuous speech = 60 ÷ L_stream, which is 60 ÷ 5 = 12 for about 5-minute streams. If F devices stream continuously and their rollovers spread evenly, the average load is F ÷ 5 requests per minute, against a limit of 900 per minute. That is engineering arithmetic on the published figures, and it assumes nothing else in the project uses the request quota. Rollovers may bunch up, so test a synchronized case.

Inputs, and which are yours to replace

What are the common traps?

These follow from the published figures above. They are engineering reasoning about how the numbers interact. None describes an incident at any company.

• Idle audio on a billed-by-open-time plan. AssemblyAI bills the time the WebSocket is open. A pendant that holds a socket open all day pays for the hours without speech. The 10-hour example above shows five times the cost of the 2-hour case at the same hourly rate.
• A low per-minute price with a stream that never closes. Amazon Transcribe bills in one-second increments, and its streaming docs say to send silence as zero bytes so the stream stays real-time. The pages we read do not say whether those silent seconds are billed. If they are, the 2-hour audio figure understates the bill. Confirm it with the vendor before you budget.
• Reconnects that count as requests. Google counts each streaming session as one request against 900 per 60 seconds per project. A fleet that all rolls over its roughly 5-minute streams on the same schedule hits that project limit in bursts.
• Hard session caps that cut a sentence in half. Amazon says the session cap cannot be raised and the client must start a new session. Where the cut lands in the speech is a question for your own test, because no vendor page states how rollover affects words near the boundary.
• Quotas that count different things. AWS and Azure count concurrent streams, Deepgram counts concurrent connections, and AssemblyAI counts new streams per minute. The numbers cannot be compared line by line. Convert each to a fleet size using your own a.
• Free-tier limits copied into a launch plan. The free AssemblyAI tier allows 5 new streams per minute, and the free Azure tier allows 1 concurrent request. A prototype that works on those limits says little about a fleet.

What should you measure on your own board?

Vendor pages do not give latency for your hardware and network, and they do not describe what happens to words at a stream rollover. We have not measured either, and no result appears here. The pass line is yours to set for each product requirement, and you record the reason behind it.

What should you think through before you choose?

Data handling is part of the choice, and we did not retrieve each vendor's retention and opt-out terms. These questions include that gap. Take the list to your team and to any supplier who proposes a design.

• Does the socket stay open while nobody is speaking, and which provider billing unit makes that expensive?
• Who decides when a stream closes: the device, the phone app or the backend?
• What does the product do at the moment a stream hits its cap or is dropped?
• Is the wearer told visibly when the device is recording and sending audio, and does the status light follow the real state of the socket?
• Does the person speaking nearby know about the recording, and how does your product collect and record consent?
• Which language list does the product need, and has it been read from each provider page again at design time?
• What does each provider keep after a stream ends, and can you opt out of training use? We did not retrieve these terms, so read each vendor data-handling page.
• Who owns the provider account and keys, and are they in your company name?
• Which quota applies to your fleet on launch day, and how many devices stream in the same minute?
• Which invoice line will you compare with your own logs each month?

Planning a recorder or pendant and want the streaming design checked against your wake-up pattern? Tell us about your device. You can also read about our AI recording wearable development and how we scope app, cloud and AI work.

Sources

All pages accessed 2026-10-31. Prices and quotas change, so confirm figures on the live pages.

• AssemblyAI, Pricing: https://www.assemblyai.com/pricing (accessed 2026-10-31)
• Amazon Web Services, Amazon Transcribe pricing: https://aws.amazon.com/transcribe/pricing/ (accessed 2026-10-31)
• Amazon Web Services, Amazon Transcribe endpoints and quotas: https://docs.aws.amazon.com/general/latest/gr/transcribe.html (accessed 2026-10-31)
• Amazon Web Services, Amazon Transcribe streaming documentation: https://docs.aws.amazon.com/transcribe/latest/dg/streaming.html (accessed 2026-10-31)
• Google Cloud, Speech-to-Text quotas and limits: https://docs.cloud.google.com/speech-to-text/quotas (accessed 2026-10-31)
• Microsoft, Azure Speech service quotas and limits: https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-services-quotas-and-limits (accessed 2026-10-31)
• Deepgram, Pricing: https://deepgram.com/pricing (accessed 2026-10-31)