AssemblyAI-TTS-1: $15 per 1M characters for voice agents

AssemblyAI has released AssemblyAI-TTS-1, its first text-to-speech model, now generally available to developers building voice agents for live phone calls. The company prices it at $15 per 1M input characters and offers unlimited concurrent requests without contracts or minimums, positioning it as the lowest-cost frontier text-to-speech model for voice agents with this concurrency model.

What changed

AssemblyAI announced the general availability of AssemblyAI-TTS-1 on October 27, 2026. The model targets voice agents handling real-time phone conversations. The company identifies two primary challenges in production voice agents: callers abandon systems that sound robotic, and text-to-speech costs at scale undermine the business case for deploying agents.

Pricing is $15 per 1M input characters, counting spaces and punctuation. Voice cloning, word timestamps, and EU region processing all operate under the same rate structure. No per-voice fees, concurrency charges, or monthly minimums apply. This approach contrasts with competitors like Inworld TTS-2 Flash, which charges $15 per 1M on-demand but restricts that tier to 5 concurrent requests. Paid tiers at Inworld drop to $7 per 1M characters.

The API includes a library exceeding 4,000 voices and instant voice cloning functionality. Voice cloning requires a short audio recording plus confirmation of speaker permission. Six languages launch at availability: English, Spanish, French, German, Italian, and Portuguese. Developers can integrate via HTTP API, WebSocket API, or an OpenAI-compatible endpoint. Output supports mono PCM, mu-law, or A-law formats at 8 kHz to 48 kHz, with 24 kHz PCM as native format. Word timestamps come included.

Why it matters

Voice agents represent an expanding market for businesses handling phone-based customer interactions. Latency and speech quality directly affect whether callers remain on the line. Developers already paying for speech-to-text transcription now face compounding costs when scaling text-to-speech synthesis across thousands of concurrent calls.

AssemblyAI’s pricing structure removes concurrency tiers entirely. At the same $15 per 1M character rate, developers can run ten concurrent calls or ten thousand without tier upgrades. This eliminates a cost driver that forces teams to accept higher per-character rates on premium plans. For developers already invested in AssemblyAI’s speech-to-text platform, this release completes the voice loop with a single vendor, reducing integration complexity.

However, benchmark results reveal performance disparities. AssemblyAI’s self-reported testing shows 47 ms median time to first byte, outperforming competitors like Inworld TTS-2 Flash at 244 ms in the same test. Independent testing by Cekura on 27 models placed AssemblyAI-TTS-1 at 159 ms median streaming latency, 8th of 19 timed models. Inworld TTS-2 Flash scored 45 ms in that independent benchmark. Self-reported figures show AssemblyAI roughly four times faster than the independent benchmark suggests, indicating significant variation between controlled testing and real-world conditions.

On quality metrics, AssemblyAI reports 93.7% entity accuracy (numbers, codes, abbreviations) versus ElevenLabs v4 at 93.6%, with predicted mean opinion score of 3.85 versus Inworld TTS-2 Flash at 3.88. These margins suggest comparable rather than superior voice naturalness.

What to test

Before adopting AssemblyAI-TTS-1 for production voice agents, verify these claims against your use case:

• Latency on live call traffic. AssemblyAI quotes 90 ms time to first audio in production as a conservative planning figure. The gap between self-reported 47 ms and Cekura’s independent 159 ms median is substantial. Test streaming latency on your actual network paths and regions (us-east-1 or eu-north-1 only) before committing.

• Entity accuracy with your domain vocabulary. The company tested on 580 hard targets like phone numbers and codes, achieving 93.7% accuracy. Confirm performance on domain-specific terms, abbreviations, and critical phrases specific to your industry.

• Voice quality impact on call completion. The predicted mean opinion score of 3.85 trails Inworld at 3.88. Conduct A/B tests with real callers if call completion rates directly impact your business outcome.

• Cost relative to your concurrency profile. Unlimited concurrency saves money only if you regularly exceed 5 concurrent requests. Calculate your monthly character volume and compare against competitors on both on-demand and paid tiers.

• Feature gaps. AssemblyAI-TTS-1 launches without SSML, speed control, pronunciation dictionaries, or emotion tags. Identify whether these features are required for your use case before adoption, or confirm roadmap timing.

• Regional constraints. AssemblyAI operates TTS-1 only in US (us-east-1) and EU (eu-north-1). Verify whether your deployment geography falls within these regions, as no on-premises or additional regional options exist.

Common questions

How much does AssemblyAI text-to-speech cost compared to competitors?

AssemblyAI-TTS-1 costs $15 per 1M characters with unlimited concurrency and no contract. Inworld TTS-2 Flash also lists $15 per 1M on demand but restricts to 5 concurrent requests. Paid Inworld plans cost $7 per 1M. The primary difference is that AssemblyAI charges the same per-character rate regardless of concurrent session count.

What latency should I expect from AssemblyAI-TTS-1?

AssemblyAI reports 47 ms median time to first byte in self-reported testing and cites 90 ms time to first audio in production as a conservative estimate. Independent benchmarking by Cekura measured 159 ms median streaming latency, so real-world results will vary based on network, region, and request size.

What languages does AssemblyAI-TTS-1 support?

Six languages launch at general availability: English, Spanish, French, German, Italian, and Portuguese.

The conclusion

AssemblyAI-TTS-1 pricing of $15 per 1M characters with unlimited concurrency directly addresses a real cost constraint for developers scaling voice agents. The removal of per-concurrency fees is a genuine advantage over competitors who tier pricing by session count.

Performance claims warrant independent verification. AssemblyAI’s self-reported 47 ms latency substantially outpaces independent benchmarks. If sub-100ms latency is critical to your caller experience, test on representative traffic before full deployment.

The model’s entity accuracy at 93.7% and predicted mean opinion score at 3.85 are competitive but not standout. Teams already satisfied with ElevenLabs or Inworld voice quality will see these as equivalent rather than superior. The primary value proposition is cost and concurrency flexibility rather than audio quality superiority.

The feature roadmap matters for long-term adoption. Absence of SSML, pronunciation control, and emotion tags limits fine-grained voice customization. Teams in healthcare, customer service, or regulated industries where tone and clarity affect compliance outcomes should clarify roadmap timing for these capabilities.

For teams already using AssemblyAI’s speech-to-text platform, this launch simplifies architecture by consolidating transcription, speech understanding, language model inference, and synthesis with one vendor. For teams established with other providers, the switching cost of API migration and latency testing may outweigh pricing savings unless concurrency fees are your primary cost driver.

Related reading

• OpenMobile-2 smartphone agents match proprietary rivals
• Agensh: Decentralized multi-agent harness for 1,024 agents
• Block Dessa speech model processes raw audio to detect emotions