Evidence at a glance
Filling the Speaking Side of a Phone Call
AssemblyAI has released AssemblyAI-TTS-1, its first text-to-speech model, for voice agents that need to speak replies during live phone calls. It is a hosted API rather than a downloadable model. Developers can call it over HTTP, WebSocket, or an OpenAI-compatible route, and it runs in AssemblyAI’s US or EU cloud regions.
The release matters to technical leads not simply because it adds another voice, but because it puts speech output alongside AssemblyAI’s existing speech products. For teams already using its speech recognition, a single vendor may shorten the integration path. Whether the model works well on calls, however, depends on when callers hear it, whether it reads the content correctly, and whether it stops promptly when they interrupt.
A First Byte Is Not the Whole Call
Real-time speech latency can be measured from different starting points. AssemblyAI reports a 47 ms median time to first byte in its own test. Cekura reports 159 ms for streaming latency and a 171 ms P50 time to first audio. These figures use different metrics and test conditions, so they are not direct contradictions—and 47 ms should not be read as the time from a caller finishing a sentence to hearing a complete reply.
Accuracy results have similar limits. AssemblyAI reports 93.7% entity accuracy across 580 difficult targets such as phone numbers and codes, just above the 93.6% it reports for ElevenLabs v4 in its test; it also reports 99.0% word accuracy. Cekura, by contrast, gives TTS-1 81.5% reading accuracy and 159 ms streaming latency, compared with 93.2% and 45 ms for Inworld TTS-2 Flash. The evidence does not establish that TTS-1 is universally faster or more accurate. It makes the test methodology—and performance on a team’s own call scripts and system path—the first things to examine.
Interfaces Built Around Interruptions and Telephony
The WebSocket interface keeps a connection open for a conversation and accepts cancellation messages. That gives an agent a way to stop an unfinished reply when a caller interrupts. Word-level timestamps can also record which part of the reply the caller actually heard. AssemblyAI has published reference agents for LiveKit and Pipecat, offering teams a starting point in familiar voice-agent frameworks rather than requiring them to build every call-control component from scratch.
The output options are also aimed at telephony: mono PCM, mu-law, and A-law are supported at sample rates from 8 to 48 kHz. The native format is 24 kHz PCM, while phone systems can request 8 kHz mu-law directly. At launch, the model supports English, Spanish, French, German, Italian, and Portuguese. The materials provide no independent quality evaluations for each language, so support describes the available interface coverage, not verified performance on accents, proper names, or real calls in every language.
Put the Price Advantage on the Same Bill
TTS-1 is priced at $15 per million characters. Its pay-as-you-go plan has no concurrency tier and does not require an annual contract. That is substantially below the listed prices for ElevenLabs Flash v2.5 at $40 and v4 at $80 per million characters. But Inworld TTS-2 Flash is also $15 on demand. For procurement, the case is therefore not that TTS-1 is cheaper than every alternative; it is whether its price, concurrency plan, and fit with an existing AssemblyAI speech-recognition setup make the operating bill simpler.
“Unlimited concurrency” is AssemblyAI’s description of its plan, not an independent guarantee of capacity under every load or service condition. Cost estimates should also account for billable characters such as spaces and punctuation, call duration, and the actual volume of agent replies. If a team needs precise pronunciation, emotional expression, or speed adjustment, a low API price may be offset by additional text processing or external components.
Production Fit Depends on Control
At launch, TTS-1 does not support SSML, speed controls, emotion tags, or custom pronunciation dictionaries. Punctuation and written fillers still shape delivery, but teams have no dedicated parameters for fine-grained voice control. Custom dictionaries and emotion tags have been announced, yet the materials give no release date. The model is also cloud-only, with no open weights or on-premises option, which matters to teams with data-residency or runtime-environment requirements.
Teams already using AssemblyAI for speech recognition can consider TTS-1 as a quick way to add speech output, then test interruption handling, proper-name pronunciation, and end-to-end time to first audio in real calls. Teams building a new voice agent should compare it with alternatives using the same prompts, region, voice, and network conditions, and establish whether other components can supply the controls they need. The current evidence puts TTS-1 on a trial shortlist; it does not justify choosing it without validation.