Homo Sapien key stats
- Latency (measured)
- —Not measured: no publicly reachable API at benchmark time. The Index never shows vendor latency estimates.
Background
Human is the baseline reference on the Humanness Index rather than a text to speech model. The four source voices on the Index, Clara, Emma, Godfrey, and Nelliot, are real people, and each of them recorded the same 20 customer support lines that every TTS model reads, false starts and filler words included. In a battle a Human recording goes head to head against a cloned TTS version of the same voice reading the same line, so the listener is answering one question: which one is human?
Why it anchors 100
Human is pinned to a Humanness score of 100, and every model is then read as a share of that human mark. It is a reference, not a competitor, so it sits outside the model and provider counts, it does not appear in the most human highlight cards, and it carries no latency or price. The recordings are loudness normalized and converted to MP3 on ingest, the same handling the synthetic clips get, so the only thing a listener compares is the voice.
Frequently asked questions
- Is Human a text to speech model?
- No. Human is real people reading the lines. It is the four source voices recorded by the actors behind them, and it exists to anchor the top of the scale so every model score reads as how close that model gets to a real person.
- How is the Human baseline recorded and scored?
- The same four actors read all 20 lines exactly as written in a quiet room, one line per file. The recordings are loudness normalized and converted to MP3, then battle head to head against the cloned TTS of the same voice. Human is pinned to a Humanness score of 100 and the rest of the field is normalized against that anchor.
Find the most human-sounding voice for your agent.
Compare the models in blind tests, read the methodology, or get in touch.
Build a TTS model? Add yours to the Index.
