WorkflowOperators

Pronunciation validator

The validator that checks each word sound by sound — phoneme-level, not word-level.

Family: Operator · type: validator_pronunciation · canvas name “Pronunciation” · beta

The Pronunciation validator answers: was each word actually pronounced correctly? It listens to each generated line and scores pronunciation phoneme by phoneme, producing a phoneme error rate (PER) and a 0–100 score from it. Lines below the bar are retried; lines at or above it pass.

This is a finer check than Accuracy. Accuracy transcribes the audio and compares words — a name said with the wrong stress usually still transcribes correctly and slips through. This validator hears the sounds themselves.

New to validators? Start with how validators work — the pass/fail routing and retry behavior here are shared by every validator.

English only. This check scores US and British English (en-us, en-gb) audio. A Generator branch in any other language is rejected — the whole pronunciation chain is an English-branch feature today.

Config

Config keyDefaultRange
threshold990–100 (score required to pass)
max_retries31–50 (regeneration attempts for a failing line)
k1.0>0–10 (PER sensitivity)
node["config"] = {"threshold": 99.0, "max_retries": 3, "k": 1.0}

The score is 100 - k * PER * 100. At the default k of 1.0 that is simply 100 minus the percentage of phonemes wrong. Raise k to widen the penalty per error.

Why the default is 99, not 70. Its sibling validators score a matter of degree, so a 70 bar makes sense there. Here, at k = 1.0, a threshold of 90 still tolerates one sound in ten. Mispronunciation is a hard defect rather than a gradient, so the default is the near-zero-defect cutoff — and a looser bar is an explicit choice you make per node.

Where it goes

Place it last among a branch’s validators. Cheaper checks — Naturalness, Noise, Accuracy — belong before it, so it only grades lines they already passed.

Its ports carry two extra options the other validators don’t have:

  • pass — must reach Export over non-fail edges, including lines that exhausted their retries and left as “gave up”. It may also feed a Phoneme Injector or a Pronunciation Corrector, because a passing line can still carry individual mispronounced words — routing is on the score, independent of the per-word findings.
  • fail — may loop back to the Generator to regenerate, or feed a Phoneme Injector, or feed a Pronunciation Corrector. The Generator and the Corrector are both retry-loop targets: their output can loop back here to be re-graded within the retry budget.
   Generator ──▶ Phoneme Injector ──▶ Pronunciation ──pass──▶ Export
                                            │
                                          fail
                                            ▼
                                  Pronunciation Corrector
                                            │
                                            └──▶ back to Pronunciation (re-grade)

A Phoneme Injector wired into this check supplies the reference pronunciation it grades against, and must sit immediately before it.

Inspect this node

client.nodes.get_node_detail(node_type="validator_pronunciation").data