Reading a result
A result is a JSON document. Check reliable before using the scores, and the scores before the words.
A real result, shortened
{
"score": 36.2,
"scores": { "total": 36.2, "accuracy": 36.6, "fluency": 55.0, "prosodic": 45.9 },
"reliable": false,
"reasons": ["not said: good"],
"differences": {
"errors": [
{ "word": "good", "position": 4, "expected": "ɡʊd", "actual": "dɪ", "confidence": 0.879, "missing": true, "phones": [...] },
{ "word": "designer", "position": 5, "expected": "dɪzaɪnɚ", "actual": "vɛləpɚ", "confidence": 0.886,
"phones": [ { "expected": "d", "heard": "v", "confidence": 0.643, "start": 1.02, "end": 1.04 }, ... ] }
],
"alignment": { "sentence_probability": 0.078, "reliable": false, ... }
},
"audio_quality": { "duration": 2.926, "level_db": -43.1, "snr_db": 31.3, "issues": [] }
}Read speech
| Field | Meaning |
|---|---|
score | 0 to 100, the total of scores |
scores | total, accuracy, fluency, prosodic, ten times the raters' 0 to 10 scale |
reliable, reasons | False when the recording does not match the sentence, when words were not said, when much was said in excess or when the audio is unfit. The scores are still there; do not show them as they are. |
transcribe | What the word recognizer heard |
differences.errors[] | One entry per word that went wrong: word, position, expected and actual sounds in IPA, confidence (probability the raters would reject it), phones[] with each sound's expected, heard, confidence, start, end, and missing when the word was not said |
differences.words[] | The same for every word, flagged or not, with its time span and mispronounced |
differences.extra[] | Speech that is not in the sentence: after, before (positions around it), start, end, heard |
differences.alignment | Whether the recording is a reading of the sentence at all: sentence_probability, reliable |
differences.heard_phones | Every sound recognised in the audio |
audio_quality | duration, level_db, clipping_ratio, snr_db and issues: too_short, silent, too_quiet, clipped, noisy |
prosody | Pitch and loudness curves, ready to draw |
formula_score, acoustic_distance | The formula of versions before 0.4 and its distance to a synthetic native reading, kept for comparison |
Free speech
| Field | Meaning |
|---|---|
mode | "free" |
score, scores | The same four dimensions, averaged over the segments by duration |
transcribe | The transcript everything is measured against |
words[] | Every word with start, end (seconds in the whole recording), confidence that it was unclear, unclear, expected and heard |
unclear_words | The words above the threshold |
segments[] | The parts cut at the pauses, each with start, end, transcribe, scores |
duration, speech_duration | Seconds |
audio_quality, reliable, reasons | As for read speech; reliable is false when there is too little speech |
Showing it to your users
Turn the total into a band rather than a decimal: two scores a few points apart are the same performance. Bands that read well: 90 and above excellent, 70 to 89 good, 50 to 69 understandable with clear errors, under 50 hard to follow. Then list at most three flagged words, with a button that replays each one from its start to its end.