FULL STORY

AssemblyAI TTS-1: From Launch to Benchmarks

AssemblyAI launched its first TTS model TTS-1 on Oct 28, targeting real-time voice agents. Follow-up benchmarks found it fastest in class but with detail shortcomings.

2026-10-28 ~ 2026-10-29 · 2 episodes · 11 posts

Episode 1 · AssemblyAI Launches Its First TTS Model, TTS-1 (2026-10-28, 5 posts)

On October 28, AssemblyAI officially released its first text-to-speech model, AssemblyAI-TTS-1, positioned for real-time conversational voice agents and available the same day. Official specs include roughly 90ms first-audio latency (a repost alternatively cited 80ms first-byte/TTFB—slightly different framings), pricing of $15 per million characters, instant voice cloning with 4,000+ preset voices, and streaming latency on the order of 90 milliseconds. The company claims no other TTS model simultaneously beats it on accuracy, naturalness, and cost, making it the best TTS for running at scale.

Confirmed

  • Release and date: announced by AssemblyAI's official account on October 28; TTS-1 available immediately
  • Pricing: $15 per million characters
  • Latency: officially 90ms first audio / streaming output (a repost cited 80ms first byte)
  • Features: 4,000+ preset voices, instant voice cloning, built for real-time voice agents
  • Third-party evaluation: Artificial Analysis benchmarked it across multiple arenas—generation speed of 162 characters/second ranks near the top; Provider Voice Arena Elo of 1,227 (1,626 appearances), placing in the top twelve of both the Provider Voice Arena and Controlled Voice Arena

Unconfirmed

  • The official post and repost differ on latency: 80ms (first byte) vs. 90ms (first audio) may reflect different measurement definitions; awaiting official clarification

Why it matters

TTS-1 is AssemblyAI's first product extending from speech recognition/understanding into speech synthesis, targeting the hot voice agent market with a low-latency, low-cost combo. Artificial Analysis's independent evaluation backs up its speed and arena rankings, but the "best overall" claim is the vendor's own—real-world performance needs more head-to-head validation.

Episode 2 · AssemblyAI-TTS-1 benchmarks: top speed, mid-tier accuracy (2026-10-29, 6 posts)

Artificial Analysis ran multiple benchmark tests on AssemblyAI's first speech synthesis model, AssemblyAI-TTS-1, painting an overall picture of "top-tier speed, mediocre details."

Confirmed

  • Speed: 162 characters per second, ranking first on the TTS speed leaderboard, ahead of Cartesia Sonic (148.6), Mistral Voxtral, and others.
  • Pricing: $15 per million characters, placing it in the top tier.
  • Pronunciation robustness: An overall score of 81.7%, ranking 16th. This benchmark uses human reviewers to judge whether models correctly read four categories of difficult text; expanding abbreviations is a clear weak spot for AssemblyAI-TTS-1.
  • Controlled Voice Arena: With all models using the same cloned voice, AssemblyAI-TTS-1 ranks 10th with an Elo of 1110—6th in the knowledge-sharing category at 1139, and 10th in both the assistant and customer service categories.
  • The company also released a sample audio clip of typical customer service dialogue (covering cancellation fees and waivers) for a direct listen to its naturalness.

Why It Matters

AssemblyAI's first foray into speech synthesis lands in the top tier on speed and pricing, signaling that competition in the TTS market is shifting from pure naturalness toward latency and cost. However, its 16th-place pronunciation robustness suggests it still needs improvement on real-world customer service text like abbreviations and proper nouns—so model selection requires weighing speed against accuracy.