AssemblyAI Launches AI Voice at 90 Milliseconds to Compete with ElevenLabs and Cartesia

AssemblyAI launches its first text-to-speech model, with a claimed latency of 90 ms, 4,000 voices, and a price of $15 per million characters.

Forty-seven milliseconds to receive the first bytes of generated voice, and a price of 15 dollars per million characters. AssemblyAI, hitherto primarily known for its audio transcription models, is entering the text-to-speech market with AssemblyAI-TTS-1. The system targets phone agents and conversational assistants that need to respond quickly, correctly pronounce sensitive information, and manage numerous simultaneous conversations.

The model is commercially available, featuring over 4,000 voices, an instant voice cloning system, and support for six languages at launch. AssemblyAI also announces no contractual cap on the number of concurrent calls under pay-as-you-go billing, with no annual commitment. A First Sound Response Announced at 90 Milliseconds

For a conversational agent, the speed of text-to-speech synthesis partly determines the fluidity of the exchange. Too long a pause between the response calculated by the language model and the beginning of speech can give the impression of a rigid automated system.

AssemblyAI distinguishes between two performance measures.

In its benchmark test, conducted from a single client by directly querying the APIs of six models, AssemblyAI-TTS-1 receives the first audio byte in 47 milliseconds at median time.

The model outperforms ElevenLabs Flash v2.5, measured at 119 ms, Cartesia Sonic 3.6 at 143 ms, Inworld TTS-2 Flash at 244 ms, Inworld TTS-2 at 280 ms, and ElevenLabs v4 at 612 ms.

The commercial figure of 90 milliseconds corresponds to a more conservative estimate of the delay before audio starts in production, taking into account network conditions and processing regions.

This number does not represent the full latency of a voice agent. Speech recognition, language model reasoning, tool calls, and telephony transmission can add their own delays.

The company states it trained the model on several million hours of audio, including real conversations. Unlike certain architectures that generate speech from discrete audio tokens, AssemblyAI-TTS-1 uses continuous audio representations, presented as a way to improve quality and reduce the time required to produce the first sounds. Correctly Pronouncing Numbers, Codes, and Addresses

The quality of a synthetic voice does not depend solely on its naturalness. When an agent confirms a phone number, a booking code, or an email address, a single pronunciation error can prevent the user from completing their task.

AssemblyAI has therefore integrated specific evaluations focusing on numbers, dates, URLs, abbreviations, and codes.

On a set of 580 sentences each containing a pronunciation difficulty, AssemblyAI-TTS-1 achieves 93.7% correctly pronounced targets, compared to 93.6% for ElevenLabs v4, 82.5% for Cartesia Sonic 3.6, and 68% for ElevenLabs Flash v2.5.

On more traditional read sentences, originating notably from Seed-TTS and LibriSpeech, the model reaches 99% word-level accuracy.

These results are calculated from transcriptions of the generated audio, and not solely from human assessment. The method allows for a reproducible comparison, but remains dependent on the recognition system used to evaluate the output.

The company's protocol also includes 500 turns from recorded customer conversations, as well as a fixed set of 50 cloned voices used to compare models under similar conditions. A Natural Voice, but Not Systematically the Best

Evaluating vocal naturalness is more subjective than measuring a transcription error.

AssemblyAI notably uses predictive models of auditory preference that estimate the rating a panel of listeners might assign to an excerpt.

Its new model obtains 3.85 out of 5, ahead of Cartesia Sonic 3.6 at 3.82, ElevenLabs Flash v2.5 at 3.70, and ElevenLabs v4 at 3.67. Inworld TTS-2 Flash remains slightly ahead, with 3.88.

The model also reaches 85.5 out of 100 on TTSDS, an indicator comparing the distribution of generated prosody to that of natural recordings. The similarity between a cloned voice and its reference recording is measured at 0.64 in published tests.

These results do not point to a single, universally superior model. Depending on whether an application prioritizes expressiveness, number accuracy, cloning fidelity, or speed, the rankings can vary. Independent Evaluations Confirm Part of the Performance

AssemblyAI also relies on rankings established by external platforms.

In Cekura's benchmark, covering 104 sentences spoken by 27 models, AssemblyAI-TTS-1 is among the systems that best combine speed and accuracy. It notably comes out on top for phone numbers and codes, with a 91.7% success rate, and ranks fourth for vocal naturalness, with a rating of 4.50.

On Coval's leaderboard, the model's word error rate reaches 1.7%, 0.3 points behind the leader and slightly ahead of certain ElevenLabs and Cartesia models evaluated.

Early results from Artificial Analysis also position AssemblyAI-TTS-1 among the top ten models in its controlled-voice arena, where several systems are compared using the same cloned voices.

These external observations provide additional insight to AssemblyAI's tests, but do not replace trials on real phone conversations. User-perceived performance also depends on call context, interruptions, and the quality of the overall voice chain. Over 4,000 Voices and Cloning Achieved in Seconds

AssemblyAI's catalog includes over 4,000 voices, accessible from its interface and API.

The platform also allows the creation of a custom voice from an audio recording. Users must confirm they have permission from the individual concerned. AssemblyAI notably showcases a clone created from 24 seconds of speech.

At launch, the system supports English, Spanish, French, German, Italian, and Portuguese, with various regional accents.

Certain features remain absent from this first version. Advanced voice cloning relying on custom training is not available. The same applies to pronunciation dictionaries, dedicated emotional tags, explicit speed control, and SSML instructions.

However, the model can interpret