Why We Built SpeakEasy

I didn't set out to build a speech API company.

I was building something else entirely — a product that needed transcription. Record audio, get text back. Simple, right? The kind of feature you figure will take an afternoon to integrate and then you never think about again.

That's not what happened.

The moment I knew something was broken

I went to the OpenAI pricing page. Then Deepgram. Then Google Cloud. Then AssemblyAI. Then back to OpenAI because surely I'd read it wrong the first time.

I hadn't.

Here's the thing: I didn't just find expensive APIs. I found confusing APIs. Pricing pages that required a spreadsheet to decode. Per-minute billing that punished short audio clips. Base rates that doubled the second you enabled streaming. Features locked behind enterprise tiers. "Contact us for pricing" on things that should cost cents.

I remember sitting there thinking: this is a solved problem. Speech recognition has been around for decades. The open-source models are incredible. Why does it cost this much to send audio to an endpoint and get text back?

The pricing game nobody talks about

The more I dug, the worse it got.

Most speech API providers lead with their lowest possible rate. It's technically accurate. It's also practically useless. That rate is for batch processing, on their worst model, without any of the features you actually need.

Want real-time streaming? That's a premium. Speaker diarization? Add-on. Word-level timestamps? Different tier. Need more than 5 concurrent connections? Time for the enterprise plan.

I did the math on what a real production workload would actually cost. Not the marketing page number — the real number, with streaming, diarization, and decent concurrency.

Google Cloud: over $1.20 per hour. Deepgram: around $0.40/hr once you add the features. AssemblyAI: $0.45/hr for streaming. OpenAI: $0.36/hr and no diarization at all.

For an indie developer building a side project? These prices kill the idea before it ships. For a startup trying to stay lean? It's a line item that grows faster than your revenue.

Honestly, the pricing felt like it was designed for procurement teams, not developers. Enterprise sales cycles. Annual contracts. "Book a demo" buttons where a signup page should be.

I wasn't the target customer. And neither were most of the developers I knew.

The frustration wasn't just price

Price was the headline problem, but the developer experience was its own disaster.

Google Cloud Speech requires you to set up a GCP project, enable billing, create a service account, download a JSON key file, set an environment variable, install the SDK, and then — finally — you can make your first API call. I counted the steps once. It was fourteen.

AWS Transcribe is worse. You need an AWS account, IAM permissions, an S3 bucket for your audio files, and the mental model of how AWS services wire together. For transcription. Of a voice memo.

Even the developer-friendlier options had friction. Deepgram's SDK is fine, but the pricing tiers are a maze. AssemblyAI is clean but expensive for what you get. ElevenLabs does great TTS but doesn't do STT, and their pricing is built for consumers, not developers.

I kept thinking: someone should just make a dead-simple API. One endpoint. Send audio, get text. Send text, get audio. Charge a fair price. Include everything. Done.

Then I realized that someone was going to have to be me.

What "affordable" actually means

I want to be specific here, because "affordable" is the kind of word that means nothing if you don't put numbers behind it.

SpeakEasy costs $10 a month. Your first month is $1. That gets you 50 hours of speech processing — STT or TTS, your choice. All features included. Diarization, timestamps, streaming, 99+ languages. No add-ons, no tiers, no "contact sales."

If you go over 50 hours, overage is $0.25 per hour. That's it. The effective STT rate is $0.20 per hour. TTS is $2.50 per million characters.

And here's something that sounds small but matters a lot: per-second billing. Not per-minute. Not per-15-second block. Per-second.

If you process an 11-second clip, you pay for 11 seconds. Not 60. Not 15. Eleven.

At scale, the difference between per-second and per-minute billing is enormous. If you're processing thousands of short audio clips — voice messages, customer support snippets, voice commands — per-minute billing silently inflates your costs by 30-40%. It's not a rounding error. It's a hidden tax.

We don't do that.

Built for developers who ship

The other thing that bothered me about existing APIs was the onboarding. I'd been spoiled by products like Stripe and Vercel — sign up, get a key, start building. No sales call. No 40-page quickstart guide. No SDK that requires three wrapper libraries and a configuration file.

That's what I wanted SpeakEasy to feel like.

Sign up. Get an API key. Make a request. That's the entire onboarding.

We're OpenAI SDK compatible. If you've already built something on Whisper, you can switch to SpeakEasy by changing one line — the base URL. Same request format. Same response format. Your code doesn't change.

We're ElevenLabs SDK compatible too, for TTS. Same deal — swap the URL, keep your code.

I built this because I'm the kind of developer who evaluates an API by how fast I can go from zero to working code. If it takes more than five minutes to get a response back, something is wrong with the product, not with me.

Why now

The timing matters.

Open-source speech models have gotten incredibly good. Whisper changed the game. The models that power transcription and synthesis today are better and cheaper to run than anything available even two years ago.

But the pricing of hosted APIs hasn't kept up. The big players are still charging 2018 prices for 2026 infrastructure costs. The margins are enormous, and they're protected by complexity — confusing pricing pages, enterprise-only features, long integration cycles.

That gap — between what speech processing actually costs to run and what developers are being charged — is where SpeakEasy lives.

I'm not claiming we've invented new AI. We haven't. I'm saying the API layer between great models and developers is broken, and we're fixing it. Better pricing. Better DX. Fewer gotchas. That's the whole pitch.

The challenger mindset

SpeakEasy is a small company going up against OpenAI, Google, Amazon, and a handful of well-funded startups. I know that. I'm not pretending otherwise.

But here's what I've learned from building products: most developers don't need the biggest provider. They need the one that respects their time and their budget. They need the one that works on a Saturday afternoon when they're hacking on a side project, not the one that requires a procurement process.

The big players are optimized for enterprise deals. They're optimized for annual contracts and dedicated account managers and SOC 2 compliance calls.

I'm optimized for developers who want to ship something this weekend.

That's a different market. And it's a market that's been underserved for too long.

What's next

We're early. The product is live. The API works. Developers are using it.

But this is just the beginning. I'm building SpeakEasy to be the default speech API for developers who care about value — the Hetzner of speech processing, if you will. Not the cheapest possible option with no support, but the best value: great quality, fair pricing, and a developer experience that doesn't waste your time.

If you've been frustrated by what speech APIs cost — or by how long it takes to get one working — I built this for you.

Try it. Your first month is $1, and you get 50 hours to put it through its paces. All features included from day one. No surprises.

If it's not for you, no hard feelings. But I think once you see what $0.20 per hour with per-second billing feels like compared to what you've been paying, you won't go back.

I didn't.