Most production voice agents use the cascaded model architecture: Speech-to-text, Large Language Model, Text-to-Speech. On a per-minute call basis, the most expensive part of the cascade is TTS. Voice agent developers need TTS that sounds like a real human, responds quickly, and doesn’t eat up all their margins.
It’s very difficult for developers to compare costs across providers because each provider uses a different character to voice agent minute conversion. To make the problem worse, there are also hidden fees for production-scale concurrency. The purpose of this blog post is to make it very easy and clear to compare costs across the leading TTS providers that are actually used at scale in production voice agents.
What TTS providers do voice agent developers actually use?
There is a deluge of TTS providers. It seems like everyone with a garage in Silicon Valley is running a TTS model company. There’s everything from open-weight models like Kokoro that sound robotic to high-end proprietary models that sound like real people. The reality is that most of these models are not good enough. Voice agent developers have spent countless hours evaluating TTS in production calls. They’ve discovered that there is only a small list of TTS models that actually sound natural enough and respond with fast enough Time To First Audio (TTFA) to run in production voice agent calls. The list is AssemblyAI, ElevenLabs, Cartesia, Rime, and Fish. Outside of that list, there is very little usage in production voice agents.
The character to voice agent minute conversion
Text-to-Speech is priced per million characters, but in voice agents we care about per-minute cost. The businesses that buy voice agents think about their costs on a per-minute basis. Voice agent developers need to know their per-minute costs so they can quote their customers in the standard industry billing unit. We have to convert the per character cost to a per-minute cost.
The tricky part is that a voice agent doesn’t speak the whole time during a call. The caller takes a turn speaking, then the agent takes a turn speaking. Different TTS providers use different conversion factors for characters to minutes. For example, Cartesia uses 750 characters per minute, while ElevenLabs uses 1,000 characters per minute. If you get quotes from both providers, and each uses different conversion factors, you are comparing apples to oranges.
At AssemblyAI, we process millions of minutes of voice agent traffic every day across every voice agent use case. We have data on actual characters per minute. We know call duration because the STT WebSocket connection is open for the entire call. We also know the number of TTS characters used in the call. From our expansive data, across many different voice agent use cases, we see that 600 characters per minute is the most accurate conversion factor. In the rest of this post, we normalize TTS costs to per-minute costs using 600 characters per minute.
Concurrency: the hidden fee
When you run a voice agent at scale, you need to serve many concurrent calls. The other voice agent TTS providers require you to sign an annual contract to unlock production-scale concurrency. You get a low amount of concurrency on their self-serve tiers. As you start to get real customer volume, you quickly max out the allotted concurrency on self-serve and have to talk to sales.
The spend commitment price per concurrent voice agent call varies by TTS provider, and they don’t publicly publish the costs. However, talk to any voice agent developer with scale and they know. The average rate is about $10 in monthly committed spend per concurrent voice agent call. If you need to make 1,000 concurrent voice agent calls, then you need to commit to spending $10,000 in TTS usage for the next 12 months. If you don’t use the amount you committed to spending each month, you still have to pay up. The TTS provider pockets your unused spend. It’s a staggering hidden fee.
What’s worse, the TTS industry has made a mess of how it talks about concurrency. The other TTS providers will tell you how many concurrent API generations you get, not how many concurrent calls your voice agent can make per API generation. Every TTS provider will tell you a different number on how many concurrent calls you can make with one API generation.
For voice agent developers, it is truly a horrible experience. Your app is taking off, and suddenly you hit a ceiling on concurrency and have to schedule a meeting with a salesperson. You are forced into committing to a spend amount for your brand new app and don’t know how much you will need over the next year. It feels icky, so you go get other quotes, only to find each provider is talking about price per minute and concurrency in different terms. Engineers love meetings and opaque policies, right?
AssemblyAI gives developers unlimited concurrency. You don’t have to commit to $10 in inference spend per month for the year for every voice agent concurrent call. And you don’t need to worry about how to translate API generations to concurrent calls. We just get out of your way so you can build your product.
AssemblyAI-TTS-1 comes with unlimited concurrency and no spend commitment. Start building on pay-as-you-go pricing today.
TTS cost comparison
As of the time of writing this post, here are the published costs for TTS models used in production voice agents in price per million characters on the highest price (lowest unit cost) self-serve tier, per minute using the same conversion (600 characters per minute), approximate spend commit for each concurrent call, and time to first audio according to Coval’s TTS benchmark.
| Provider | Model | $ / 1M characters | $ / minute | Monthly spend commit / concurrent call | TTFA |
|---|---|---|---|---|---|
| AssemblyAI | AssemblyAI-TTS-1 | $15 | $0.009 | Unlimited | 90 ms |
| ElevenLabs | v4 Turbo | $40 | $0.024 | ~$10 | 190 ms |
| Cartesia | Sonic 3.6 | $37.50 | $0.0225 | ~$10 | 285 ms |
| Rime | Coda | $50 | $0.03 | Unlimited | 395 ms |
| Fish | Drama 3 | $50 | $0.03 | ~$10 | 370 ms |
Compare voices, naturalness, and responsiveness on your own text before you run the numbers on your call volume.
Try out AssemblyAI-TTS-1
AssemblyAI is the only TTS voice agent developers use in production that offers frontier naturalness, low latency (~90 ms TTFA), unlimited concurrency, at one-third the per-minute price of the other TTS providers.
Try it out! Log into our dashboard to hear the voices, build a voice agent with our LiveKit and Pipecat reference implementations, or update your voice agent code. We’d love to hear what you think of our TTS model.
Pair AssemblyAI-TTS-1 with LiveKit or Pipecat and get production-ready text-to-speech at $15 per 1M characters.
Frequently Asked Questions (FAQs)
How many characters per minute does a voice agent use for text-to-speech?
Across AssemblyAI’s voice agent traffic, a call uses about 600 TTS characters per minute, because the agent only speaks for part of each call. Provider quotes often assume more. Use one conversion factor for every quote so the per-minute numbers are comparable.
How do you calculate TTS cost per minute for a voice agent?
Multiply the price per 1M characters by 600, then divide by 1,000,000. At $15 per 1M characters, that’s $0.009 per minute; at $40 per 1M characters, it’s $0.024 per minute. Running every provider through the same factor turns per-character price cards into a like-for-like per-minute comparison.
How does TTS concurrency pricing work?
Most TTS providers cap how many requests you can run at once on their self-serve plans. To go past that cap, you have to talk to sales and sign an annual contract with a monthly spend commitment, which you pay whether or not you use it. AssemblyAI gives developers unlimited concurrency without a spend commitment.
How much does AssemblyAI text-to-speech cost?
AssemblyAI-TTS-1 costs $15 per 1M characters, billed on input characters with no promotional pricing. At 600 characters per minute, that works out to about $0.009 per voice agent minute. Cloned voices and EU processing are the same price.
Which TTS providers are used in production voice agents?
AssemblyAI, ElevenLabs, Cartesia, Rime, and Fish account for most production voice agent TTS. These are the providers whose voices sound natural enough and whose time to first audio is fast enough for live calls. Other models see very little production voice agent usage.
