speechocean762: An Open-Source Non-native English Speech Corpus For Pronunciation Assessment

1. Introduction

• Pronunciation assessment is used in language-learning applications and proficiency tests, and research interest is rising amid teacher shortages and online learning.The paper also cites evidence that CAPT can improve foreign-language pronunciation.Source: §1 · p. 1
• Sentence-, word-, and phoneme-level assessments differ in granularity, and word accuracy is not simply the average of phone-level accuracy.The authors illustrate the distinction with two single-phone errors in “above” that listeners may judge differently.Source: §1 · p. 1
• Existing resources vary in language, speaker population, annotation type, and availability, including corpora with word- or phoneme-level labels and others unavailable publicly.The related-work survey also describes non-English corpora and resources restricted to academic use.Source: §1 · p. 1
• Existing non-native English corpora do not jointly provide free commercial and non-commercial access, child and adult speakers, and manual sentence-, word-, and phoneme-level annotations.The authors identify this combination as missing from existing corpora.Source: §1 · p. 1
• The corpus was created to support pronunciation assessment research, with its baseline system available through OpenSLR and Kaldi.The paper describes the Kaldi recipe as illustrating phoneme-level assessment.Source: §1 · pp. 1–2

2. Audio Acquisition

• Each speaker read 20 sentences from everyday text containing about 2,600 common English words, producing about six hours of audio overall.Speakers held mobile phones 20 cm from their mouths and read in a quiet 3×3-meter room.Source: §2 · p. 2
• The corpus contains 250 English learners whose mother tongue is Mandarin, split randomly into training and test sets of 125 speakers each.The split is by speaker.Source: §2 · p. 2
• Speaker selection considered gender, age, and English proficiency; experts rated proficiency as good, average, or poor, and adult and child groups each had a 1:1 gender ratio.Figure 2 shows the proficiency distributions and Figure 3 the age distributions.Source: §2 · p. 2

3. Manual Annotation

• Each utterance was independently scored by five experts under the same metrics, with repeated listening and a warning to recheck apparently inconsistent scores.Experts listened at least three times and voted on canonical sequences before scoring words that had alternatives.Source: §3 · p. 2, Figure 4 · p. 2, §3.3 · p. 3

3.3. Scoring Workflow

• The SpeechOcean uTrans application supports expert scoring after canonical phone-sequence agreement and voting.For phoneme scoring, the expert selects a phone symbol and assigns 0 or 1; an unselected symbol defaults to 2.Source: Figure 4 · p. 2, §3.3 · p. 3

3.1. Manual Scoring Metrics

• Phoneme scores measure phone accuracy; word scores cover accuracy and stress; sentence scores cover accuracy, completeness, fluency, and prosody.Completeness is the percentage of target-text words actually pronounced.Source: §3.1 · p. 2, Table 1 · p. 3

3.2. The Multiple Canonical Phone Sequences Problem

• Experts voted on a canonical phone sequence for words with multiple plausible pronunciations, then all used the agreed sequence to score.This avoids judging a pronunciation against a reference sequence that does not match the speaker’s pronunciation.Source: §3.2 · pp. 2–3

3.4. Score Distribution

• Sentence-level scores ranged from 3 to 10, while most word- and phoneme-level scores fell between 8 and 10.The authors attribute this pattern to sentence scores depending on consistently good word and phone pronunciation; even one mispronounced word can lower the overall score.Source: §3.4 · pp. 3–4, Figure 7 · p. 4, Figure 8 · p. 4

4. The Kaldi Recipe

• The released “gop speechocean762” Kaldi recipe demonstrates phoneme-level scoring using a classical neural-network GOP pipeline and phone-specific regressors.It obtains frame-level posteriors from a native-speech acoustic model, then uses forced alignment to compute GOP values and features.Source: §4 · p. 4, §4.1 · p. 4, §4.3 · p. 4

4.4. Results

• The authors describe the classical NN-based GOP baseline as not quite strong; Table 2 reports its MSE and PCC against manual scores.The supplied passages do not give Table 2’s metric values.Source: Table 2 · p. 4, §4.4 · p. 4

4.2. Alignment Graph Building without Lexicon

• To avoid uncertain alignments when a lexicon contains multiple pronunciations, the recipe builds LG directly from expert-voted canonical phone sequences.The construction adds skippable silence between words and uses a disambiguation symbol at the graph tail.Source: §4.2 · p. 4

4.3. Supervised Training and Data Balancing

• The recipe trains a support vector regressor for each mono phone from GOP-based features and manual scores, with polynomial regression as a lightweight alternative.To address unbalanced phone-level scores, high-scoring samples of other phones are added as low-score examples for the current phone.Source: §4.3 · p. 4

4.4. Results

• The recipe’s predictions are evaluated against manual scores using mean squared error and Pearson correlation coefficient.The results are presented in Table 2.Source: §4.4 · p. 4, Table 2 · p. 4

• The released corpus includes child and adult speech manually annotated by five experts at sentence, word, and phoneme levels.Source: §5 · p. 4
• A Kaldi recipe illustrates the classic GOP method for phoneme-level scoring.Source: §5 · p. 4
• The recipe is slated for future expansion to word- and sentence-level scoring.Source: §5 · p. 4