How Much Are Free Speech-to-Text Credits Worth?
Free speech-to-text credits run from small rate-limited developer tiers up to grants worth $150,000, and the category is unusually easy to fund because transcription also sits inside the much larger cloud programs you may already qualify for.
Speech is one of the few AI bills that converts cleanly into a forecast. The API charges per hour of audio processed, so a balance is not an abstract allowance, it is a number of hours.
AI Perks tracks the current terms across this category alongside $7.7M in credits from 194 companies.
| Provider | What the credits buy | Tracked credit value |
|---|---|---|
| AssemblyAI | Async and streaming transcription plus the derived audio intelligence layer | Up to $150,000 |
| AWS | Amazon Transcribe drawn against general cloud credits | Into six figures on the cloud track |
| Google Cloud | Speech-to-Text alongside compute and storage on one grant | Five to six figures, tiered by stage |
| Microsoft Azure | Azure AI Speech inside the broader founders grant | Five to six figures, tiered by stage |
| ElevenLabs | Speech-to-text and text-to-speech on a single vendor bill | $5,000 |
| Deepgram | Low latency streaming and batch transcription | $200 to $2,000 tracked range |
| OpenAI | Whisper and the newer transcription endpoints | Varies by track, no single published headline |
| Groq | Hosted open-weight transcription at very high speed | Free developer tier, rate limited |
| Speechmatics | Multilingual transcription with broad accent coverage | Negotiated, no published headline figure |
Read those figures as hours of audio, not as money. A large grant on an expensive realtime tier buys fewer hours than a mid-sized one somewhere cheaper, and several of the numbers above are shared cloud balances.
What Speech-to-Text Is Actually For
A speech-to-text API turns audio into a transcript, and then into structured data derived from that transcript. The transcript is the commodity. The derived layer is what you are actually buying.
Open-weight models will produce a usable transcript of clean audio on a laptop. What separates a vendor API from a weekend project is everything wrapped around the words:
Speaker diarization. Labelling who said what. Non-negotiable for anything with more than one person in the room, and the hardest part to get right yourself.
Streaming. Partial results returned while the person is still speaking, for live captions and voice agents. A different engineering problem from batch, and priced differently.
Timestamps and word-level alignment. What makes a transcript clickable and searchable rather than a wall of text.
PII redaction. Stripping names, card numbers and health details before the transcript is stored. Founders undervalue it until a fintech or healthcare buyer raises it in a security review, and retrofitting it then costs far more than buying it up front.
Audio intelligence. Summaries, topic detection, sentiment and entity extraction built on top of the transcript.
Three product shapes dominate real usage: meeting notetakers, support and sales call analytics, and voice agents that must hear a user before they can answer.
How Speech-to-Text Costs Behave at Scale
Speech APIs bill per hour of audio processed, so the bill tracks how much your users talk, not how many of them sign up. Ten thousand dormant accounts cost nothing. Two hundred sales teams recording every call cost a great deal.
It is the opposite of the seat-based model most founders bring to pricing, which is why generic per-user projections fail badly here.
Published rates move constantly and vary by vendor, tier and language, so treat the left column as a range of shapes rather than a quote:
| Effective rate per audio hour | Hours from a $10,000 balance | Hours from a $150,000 grant |
|---|---|---|
| $0.60, premium realtime | ~16,700 | ~250,000 |
| $0.36, standard realtime | ~27,800 | ~417,000 |
| $0.25, standard async batch | ~40,000 | ~600,000 |
| $0.12, high volume async | ~83,300 | ~1,250,000 |
The gap between the top and bottom rows is 5x on identical spend, and that spread is decided by architecture rather than by negotiation.
Transcription is also one of the few AI costs you cannot engineer down by switching to a cheaper model. The audio duration is fixed. Your real levers are these:
Trim silence before you send. Recorded meetings are substantially dead air, and voice activity detection cuts billed hours with zero quality loss.
Never transcribe the same file twice. Store transcripts as durable artifacts, not as a cache.
Separate async from realtime deliberately. Use streaming only where a human is actually waiting for the words. Everything else is batch.
Sample for analytics. Scoring every call for a monthly trend report is waste. Ten percent usually produces the same insight.
How to Choose Between the Speech-to-Text Providers
Choose on audio conditions and latency needs, not on benchmark word error rate. A vendor that wins on clean read speech can lose badly on a noisy call recorded through a laptop microphone, which is the audio you will actually receive.
You need a voice agent that answers in under a second. Latency is the whole specification. Deepgram and Groq are built around speed, and 300ms versus 900ms is the difference between a conversation and an awkward pause.
You need the derived layer more than the transcript. AssemblyAI bundles summarisation, redaction and topic detection, which is the difference between shipping in a week and building a pipeline.
You already have a large cloud grant. AWS Transcribe, Google Cloud Speech-to-Text and Azure AI Speech all draw against a balance you may already hold, which makes them the cheapest option in the category by a wide margin.
Your product also speaks back. ElevenLabs covers both directions on one vendor relationship, which halves the procurement and billing surface for voice agent teams.
You are multilingual from day one. Accent and language coverage vary far more between vendors than English benchmark scores suggest.
AI Perks lists which of these currently have open programs and what each one requires.
Why the Order You Claim Credits In Changes the Total
Credit balances are not interchangeable, and the sequence you claim them in changes how many audio hours you end up with. Some balances cover transcription as one line of a much broader bill. Others cover nothing else and start depleting the moment they are accepted.
Three properties decide where any given balance belongs in the queue:
Breadth. A balance that also pays for compute, storage and egress does more work than a transcription-only balance of the same headline size. Voice products carry all three.
Clock sensitivity. Some balances sit dormant until you use them. Others deplete as soon as they land, which makes an early claim expensive when your first real cohort is still some way off.
Overlap with what you already hold. Text-to-speech, telephony and LLM balances bill separately from transcription, so they extend runway rather than duplicating it. Two balances covering the same line item extend nothing.
Small developer tiers are the clearest case. A few hundred dollars is sized for a benchmark against your own recordings, not for production, so it is worth something only in the week you actually run that benchmark.
getaiperks.com tracks which programs are currently open and what each one covers, which is what any sensible order depends on.
What Founders Get Wrong About Speech-to-Text Credits
The most expensive mistake is assuming self-hosting an open-weight model is free. It moves the cost to GPU hours plus the engineering time to keep diarization, punctuation and formatting acceptable, and below a certain volume that is more expensive, not less.
Four more that cost real money:
Applying before there is audio to spend it on. Grants burn against the calendar as well as against usage. A large allocation that lands before your first real cohort reaches its end mostly unspent.
Budgeting the transcript and forgetting the derived layer. Speaker labels, redaction and summaries each add either to the bill or to your build queue. Usually both.
Ignoring storage. Raw audio is large and nobody budgets it. Speech credits cover the transcription, never the bucket the recordings sit in.
Taking one grant and stopping. Speech is a single line of a voice product's bill, and the teams who get a year of runway out of credits treat programs as a queue rather than as one decision. That is the point of tracking all 194 of them at getaiperks.com.
Frequently Asked Questions
Which speech-to-text API gives startups the most free credits?
AssemblyAI carries the largest dedicated speech grant tracked, at up to $150,000. The hyperscaler cloud programs can be larger still, but transcription draws against a shared balance rather than a speech-only allocation. The right answer depends on your latency and language needs. Current terms are tracked at getaiperks.com.
Is self-hosting Whisper cheaper than paying a speech-to-text API?
Below moderate volume, no. You replace a per-hour bill with GPU hours you pay for whether or not audio arrives, plus the engineering time to make diarization, punctuation and timestamps production quality. Self-hosting wins at sustained high volume with predictable load, which is a later problem than most teams assume.
Can I stack speech-to-text credits with LLM credits?
Yes, and you should. They cover different bills. Speech credits pay for turning audio into text. LLM credits pay for the summaries, answers and agent behaviour built on top of that text. Holding both is how teams cover a full voice product for a year through getaiperks.com.
Do speech credits cover real-time streaming as well as batch?
Usually yes, but at a different rate. Streaming is consistently priced above async batch because it holds a connection open and returns partial results. A balance therefore buys materially fewer realtime hours than batch hours, which is why splitting the two deliberately is the highest leverage cost decision.
What does speech-to-text actually cost without credits?
Published rates commonly sit in the region of ten to sixty cents per hour of audio depending on vendor, tier and whether you need streaming. Rates move often and volume discounts matter at scale, so benchmark on your own recordings rather than on a pricing page.
How accurate are speech-to-text APIs in practice?
Far more accurate on clean studio audio than on the audio you will receive. Benchmark word error rates are measured on tidy recordings, while real products ingest laptop microphones, crosstalk and background noise. Vendor differences are small on clean audio and large on messy audio.
Let the machines do the listening, and let someone else pay for the hours.