# AssemblyAI-TTS-1: A New Text-to-Speech API Built for Live Voice Agents
## Overview
AssemblyAI has introduced its first text-to-speech model, AssemblyAI-TTS-1, now available for general use. The model is designed specifically for voice agents that handle live phone calls, where fast response times and clear pronunciation are critical to keeping callers engaged. At $15 per million characters, with no long-term contracts or paid concurrency tiers required, it aims to make high-quality synthetic speech affordable and straightforward to integrate.
## What Makes It Different
Most text-to-speech systems generate audio by stringing together discrete tokens. AssemblyAI-TTS-1 takes a different approach, working with continuous audio representations throughout the generation process. The company says this design choice leads to more natural-sounding output and faster delivery of the first spoken word.
The service comes with access to over 4,000 voices and supports instant voice cloning. To create a custom voice, you upload a short audio clip along with confirmation that the speaker has given permission. At launch, the system supports six languages: English, Spanish, French, German, Italian, and Portuguese.
## How Developers Use It
Integration is flexible, with three main options available:
– **HTTP API** — straightforward request-and-response calls
– **WebSocket API** — maintains a persistent connection per conversation, enabling real-time streaming and cancellation for interrupted speech
– **OpenAI-compatible route** — works with existing tools and frameworks built around the OpenAI API format
Audio can be delivered in mono PCM, mu-law, or A-law formats at sample rates ranging from 8 kHz to 48 kHz. Word timestamps are included with every response, which is useful for synchronizing text with audio. AssemblyAI also offers reference implementations built on LiveKit and Pipecat to help teams get started quickly.
## How It Performs
In AssemblyAI’s own internal tests, TTS-1 achieved 93.7% accuracy on entity recognition tasks involving difficult targets like phone numbers, dates, and codes. This was slightly ahead of ElevenLabs v4, which scored 93.6%, and well above competing models. The median time to deliver the first byte of audio was 47 milliseconds across six models in their test suite.
Independent evaluations told a more varied story. On the Cekura benchmark, TTS-1 recorded 159 milliseconds of median streaming latency and 81.5% reading accuracy. Inworld TTS-2 Flash outperformed it on both metrics in that same test, while also matching the same price point. AssemblyAI also noted a 1.7% word error rate on the Coval leaderboard and a top-10 debut in the Artificial Analysis controlled-voice arena.
## How It Compares to Competitors
| Feature | AssemblyAI-TTS-1 | Inworld TTS-2 Flash | ElevenLabs Flash v2.5 | Cartesia Sonic 3.6 |
|—|—|—|—|—|
| Price per 1M characters | $15 | $15 (on demand) | $40 | $37–$50 |
| Languages supported | 6 | 200+ | 32 | 44 |
| Median TTFB (internal test) | 47 ms | 244 ms | 119 ms | 143 ms |
| Cekura reading accuracy | 81.5% | 93.2% | 64.4% | 74.9% |
| Cekura naturalness | 4.50 | 4.46 | 4.33 | 4.41 |
| Voice cloning | Instant only | Instant and professional | Yes | Instant and professional |
| Concurrency | Unlimited | 5–500 by plan | Not disclosed | 2–15 by plan |
The standout advantage of AssemblyAI-TTS-1 is the combination of competitive pricing and unlimited concurrency. Inworld and Cartesia both charge $15 or more, but limit concurrent requests based on the subscription plan. AssemblyAI does not.
## What’s Missing at Launch
While the model performs well on core speech quality, several features are not yet available:
– **SSML support** — no markup language for fine-tuning speech output
– **Speed and pitch control** — no parameters to adjust delivery pace or tone
– **Pronunciation dictionaries** — custom word pronunciations are not yet supported (though they are announced for future release)
– **Emotion tags** — developers cannot currently tag text with emotional intent
– **Pro voice cloning** — only instant cloning is available; fine-tuned professional cloning is not yet offered
## Pricing Details
The cost is $15 per million input characters, and this flat rate applies to every voice, language, feature, and region. There are no additional per-voice fees, no minimum volume commitments, and no concurrency surcharges. Processing in the EU (eu-north-1) and the US (us-east-1) are both priced the same. This makes it easy to estimate monthly costs without worrying about hidden fees or overage charges.
## FAQ
**What is AssemblyAI-TTS-1?**
It is a streaming text-to-speech model from AssemblyAI designed to convert agent responses into spoken audio for live phone calls and voice agent applications.
**How much does it cost?**
It costs $15 per 1 million characters, with no annual contract, no minimum volume, and no additional charges for concurrency or EU processing.
**Which languages does it support?**
Six languages are available at launch: English, Spanish, French, German, Italian, and Portuguese.
**How fast is it?**
AssemblyAI reports a median time to first byte of 47 ms in its own tests. Independent benchmarks show higher latency, at around 159 ms median streaming latency on the Cekura benchmark.
**Can I clone a custom voice?**
Yes. Instant voice cloning is available, requiring only a short audio clip and the speaker’s permission.
**Does it support SSML or emotion control?**
No, these features are not available at launch but are planned for future updates.
**Can I run it on my own servers?**
No, AssemblyAI-TTS-1 is a hosted cloud API available only in the US and EU. There is no on-premise or open-weight version at this time.
**What output formats are supported?**
Mono PCM, mu-law, and A-law audio at sample rates between 8 kHz and 48 kHz.
## Conclusion
AssemblyAI-TTS-1 enters the text-to-speech market with a clear focus: affordable, high-quality voice synthesis for live agent applications. Its combination of low pricing, unlimited concurrency, and solid benchmark performance makes it a compelling choice for developers building voice-driven products. The absence of SSML, emotion tags, and professional voice cloning are notable gaps, but the company has indicated these features are on the roadmap. For teams already working with AssemblyAI’s speech-to-text tools, TTS-1 completes the voice loop under one roof, simplifying the architecture and reducing vendor complexity. As the voice agent space continues to grow, a focused and fairly priced option like this is worth watching closely.
Thank you for reading
