AssemblyAI's TTS-1 Beats ElevenLabs on Speed With 47ms First Audio

AssemblyAI has expanded beyond speech-to-text with AssemblyAI-TTS-1, its first text-to-speech model. The API targets real-time voice agents, where delayed responses create awkward pauses and pronunciation errors can derail phone numbers, confirmation codes, URLs, and medical terms.

AssemblyAI prices the model at $15 per million characters. That rate includes every voice and supported language, instant voice cloning, word-level timestamps, and US or EU data residency. The company lists no minimum spend, per-voice charge, concurrency fee, or separate feature tier.

47 ms to first audio

In AssemblyAI’s six-model test, TTS-1 began streaming audio in a median 47 milliseconds. The next-fastest model took 119 milliseconds, while the slowest took 612 milliseconds. Time to first audio measures the delay before speech begins, a critical part of perceived responsiveness in a live conversation.

AssemblyAI attributes the low latency to continuous audio representations. Many text-to-speech systems first generate discrete audio tokens and then decode them into sound. TTS-1 avoids that separate tokenization stage, reducing the work required before it can start returning audio.

The company trained the model on millions of hours of audio, including conversational recordings. That training mix is intended to improve delivery on call-like text, which often contains fragments, identifiers, abbreviations, and less formal phrasing than read-aloud datasets.

Pronunciation drives the benchmark lead

AssemblyAI compared TTS-1 with Cartesia Sonic 3.6, Inworld TTS-2, Inworld TTS-2 Flash, ElevenLabs v4, and ElevenLabs Flash v2.5. Its pronunciation test included phone numbers, dates, URLs, confirmation codes, and drug names.

Predicted mean opinion score is an automated estimate of how natural and pleasant speech sounds. AssemblyAI’s score trailed Inworld TTS-2 Flash by 0.03 and exceeded the reported scores for Cartesia Sonic 3.6 and ElevenLabs v4.

Two third-party leaderboards produced similar results:

• Cekura placed TTS-1 in its “Best Quadrant” for accuracy and latency, ranked it first on phone numbers and codes at 91.7%, and ranked it fourth on naturalness at 4.50.
• Coval measured a 1.7% word error rate, compared with 1.4% for the leader and 1.8% for ElevenLabs v4 Turbo and Cartesia Sonic 3.6. Lower word error rates indicate more accurate output.

Latency and quality scores vary with text length, network location, audio settings, and workload. Teams evaluating the API should repeat tests with their own prompts, regions, concurrency levels, and interruption patterns.

Concurrency stays off the invoice

AssemblyAI applies the same per-character rate across usage levels and advertises unlimited concurrency without a negotiated tier. A deployment handling ten simultaneous calls and one handling thousands use the same API and pricing model, subject to the service’s operational capacity and standard account controls.

The policy removes a common procurement step for voice-agent teams, which often must negotiate higher concurrency limits as call volume grows. It also makes projected text-to-speech costs easier to calculate from character volume.

4,000 voices, with launch limits

The initial catalog contains more than 4,000 voices and includes instant voice cloning. AssemblyAI says the clone demonstrated in its launch post used 24 seconds of source audio. Launch languages include English, Spanish, French, German, Italian, and Portuguese.

Several controls remain unavailable:

• Pro, or fine-tuned, voice cloning is not yet supported.
• Custom pronunciation dictionaries and emotion tags are planned but unavailable at launch.
• The API lacks SSML and an explicit speech-speed parameter.

Teams can influence delivery by having an upstream language model add punctuation and conversational disfluencies to the generated text. Explicit controls over speed, pronunciation, and emotion will still require the planned API features.

A TTS-stage swap for agent frameworks

AssemblyAI published reference implementations for a LiveKit agent and a Pipecat agent. Both show how to replace the text-to-speech stage while retaining the surrounding transcription, language-model, and turn-management pipeline.

The API supports streaming output, word-level timestamps, and a cancellation message for barge-in, allowing an agent to stop speaking when a caller interrupts. Production evaluations should cover cancellation speed, timestamp accuracy, regional latency, voice consistency, and behavior under peak concurrency.

A lower floor for TTS pricing

AssemblyAI’s $15-per-million-character rate undercuts listed prices of $40 for ElevenLabs and $30 for Deepgram frontier models. Provider discounts, character accounting, model tiers, and audio quality can change the final comparison, but the launch gives high-volume voice-agent teams another option with predictable pricing and no separate concurrency negotiation.