Reference

Glossary

Plain-language definitions of the voice and TTS terms that matter — the ones you’ll meet most.

The voice field has a lot of shorthand. Here are the terms you’ll meet most, defined plainly — with a pointer to where each one lives in Onepin when it does.

The pipeline

TTS (Text-to-Speech) — Turning written text into spoken audio. The core of every run: the Generator calls a TTS model to voice each line.

STT (Speech-to-Text) — The reverse of TTS: transcribing spoken audio back into text. Onepin uses STT inside the Accuracy validator to check that the generated audio actually said the script.

Text normalization — Rewriting text the way it’s spoken before generation — “$5” becomes “five dollars”, “Dr.” becomes “Doctor”. The Normalizer does this so numbers and symbols don’t get read literally.

voice_map — Onepin’s per-language assignment of a voice and model inside the Generator. Each language maps to one or more voices.

Provider — The company behind a voice model — ElevenLabs, Cartesia, Naver, and others. Onepin routes to providers; you pick which voice each line uses. See Voices & Models.

Native model — A provider’s own model ID, like eleven_multilingual_v2 — as opposed to the Onepin catalog ID. voice_map takes the native model ID, which is a common identifier trap.

Line — One unit of script — a sentence or a row — that flows through the graph and becomes one piece of audio. Sometimes called a take.

Locale / BCP-47 — The standard code for a language and region, like en-us or ko-kr. Onepin uses BCP-47 anywhere you set a language.

SSML — Speech Synthesis Markup Language: tags that tell a TTS model how to speak — pauses, emphasis, pronunciation. Some providers accept inline markup in the script.

Quality and validation

WER (Word Error Rate) — How far a transcript strays from the intended words, as a percentage — the insertions, deletions, and substitutions over the total. Lower is better. Onepin reports the flip side, word accuracy, as the Accuracy validator’s score.

MOS (Mean Opinion Score) — A 1–5 rating of audio quality averaged across listeners — the long-standing yardstick for how good speech sounds. A naturalness score is a machine estimate of the same idea.

Naturalness — How human the speech sounds — its rhythm, stress, and flow — rather than whether the words are right. The Naturalness validator scores it 0–100.

Prosody — The melody of speech: pitch, pace, stress, and pauses. Good prosody is what separates a flat read from a natural one.

Phoneme — The smallest unit of sound in a language — the “k” in “cat”. TTS models build speech from phonemes, and pronunciation dictionaries steer them.

Threshold — The score a line must reach to pass a validator. You set the bar; Onepin holds it.

Retry — Regenerating a line that failed a validator, up to a set budget, then delivering the best attempt. See validators.