Como o HolosAccent transforma sua voz em uma pontuação de pronúncia

Preparing your recording

First, the audio is decoded into a standard format: mono sound sampled at 16 kHz.

The system estimates the recording’s signal-to-noise ratio. If that estimate falls below 25 dB and noise filtering is available, it applies a noise-reduction filter before analysis. If filtering fails, it continues with the original recording.

The waveform is then normalized to match the speech analysis model’s input requirements. Silence and recordings shorter than half a second receive a zero score.

Allowing different pronunciations

Next, the reference text is converted into phonetic pronunciations using a pronunciation engine. A word can have multiple accepted forms.

The pronunciation set includes:

• British and American variants.
• Common reduced forms, such as the unstressed vowel in to.
• Selected contextual alternatives.

For example, the final consonant in close the door differs from the one in a close friend. The system can choose among the supported alternatives as it processes the sentence.

Listening to sounds, not reconstructing words

The recording passes through an acoustic model. It produces evidence for different speech sounds across time, at roughly 20-millisecond intervals.

This is a phonetic representation of the recording. The scorer does not use a word-level language model to reconstruct the sentence.

The system then examines those acoustic outputs in two ways:

Método	O que faz
Expected-sound alignment	Finds the best timing for the expected sounds, choosing a complete pronunciation variant for each word.
Independent sound recognition	Follows the sounds favored by the model without requiring the reference sentence.

Both views are necessary because alignment can place an expected sound somewhere in a recording even when the speaker omitted it. Independent sound recognition supplies additional evidence about what was actually spoken.

Matching also takes timing into account, making it harder for a sound in one word to count as evidence for a distant word.

Scoring each sound

For each aligned sound, the scorer calculates a pronunciation accuracy measure: how strongly the model supports the expected sound relative to competing sounds during that interval.

It also considers:

• The expected sound’s peak probability.
• Whether independent sound recognition found a matching sound nearby.

A fitted mathematical mapping turns this evidence into a sound accuracy value between 0 and 100.

A sound missing from the independent recognition result does not automatically receive zero. The acoustic evidence still matters.

Deciding which words need attention

Word feedback combines these sound values with explicit error checks. A word must:

• Have enough evidence that it was spoken.
• Meet the configured average accuracy threshold.
• Avoid detected substitutions, insertions, or severe missing-sound errors.

Clear substitutions require weak support for the expected sound and strong support for a competing sound.

Evidência	Limite atual
Support for the expected sound	Below 0.05
Support for its replacement	At least 0.5

These are model evidence thresholds, not probabilities that a human listener would judge the word incorrect.

For flagged words, the system returns the expected pronunciation and its estimate of the sounds heard, written in the International Phonetic Alphabet (IPA). That estimate can be imperfect, especially around word boundaries.

Building the sentence score

The sentence score starts with a fitted combination of:

• Average sound accuracy.
• The share of expected sounds matched by independent sound recognition.
• The number of speech sounds pronounced per second.
• Long pauses between words.

The timing calculation excludes silence before the first aligned sound and after the last.

These features contribute to one overall score. The current implementation does not provide separate user-facing fluency, completeness, or intonation scores.

How word credit affects the result

The scorer then applies an explicit word-credit multiplier. Each word is weighted by the number of sounds in its selected pronunciation.

Resultado da palavra	Crédito ganho
Correct	Full weight
Spoken but flagged	One quarter of its weight
Omitted	No credit

The calculation is:

word credit =
    (correct sound weight + 0.25 × flagged spoken sound weight)
    / total expected sound weight

raw score =
    sentence estimate, clipped to 0–100, × word credit

Consequently, a long incorrect word usually costs more than a short one. Reading only a small part of the sentence also limits the score, even if that fragment sounds clear.

Comparing your score with the reference

Finally, the system can normalize the raw score using a reference recording of the same phrase.

It scores the synthesized reference with the same scorer version and, when a usable matching baseline is already available, applies:

displayed score = min(100, raw score / reference score × 100)

For example, a raw score of 80 against a reference score of 88 becomes approximately 90.9.

This adjustment accounts for the fact that the model may give even the reference recording less than 100.

Baselines are tied to:

• The reference voice and speech-generation model.
• The scoring model.
• The vocabulary.
• The calibration and rule version.

If a matching baseline is unavailable or unusable, the result keeps its raw score. A first attempt can therefore remain unadjusted while a baseline is prepared for later attempts.

What your score means

The resulting number is a practice score on HolosAccent’s scale.

A score of 90 does not mean that 90 percent of your words were correct, and it does not establish a CEFR level or native-speaker equivalence.

The mapping has been fitted using expert-rated recordings, but the new scoring system remains experimental and its full acceptance criteria have not yet been verified. It can miss errors or flag acceptable speech.

Individual word feedback and patterns across repeated attempts provide context for interpreting the number.