ABHammad/Vokidoo-dataset

SLP-25: synthetic child speech with speech-sound errors

A multilingual, fully synthetic benchmark of child-like speech, 868 clips in 4 languages (English, German, Spanish, Egyptian Arabic), each labelled with whether the speech is typically developing or shows a speech-sound error, and which one.

Why this dataset

Recordings of children with speech-sound disorders are scarce, sensitive and almost only available in English: collecting them needs parental consent, clinical annotation and careful protection of minors' data. This makes it hard to answer a simple question: can today's speech and audio models hear a child's speech error, in more than one language?

This dataset makes that question testable without recording a single child:

• Known ground truth. Every error is placed deliberately by fixed phonological rules, so the label of each clip is known exactly, and checked afterwards.
• Several languages, the same labels. The same categories and a similar mix in every language, so models can be compared across languages.
• No personal data. All voices are synthetic; no child was recorded.

What it lets you find out:

• Can a model hear a child's mistake at all? Voice assistants, reading apps and speech-practice apps for children are built on general speech and audio models. This set tests that ability before anyone relies on it with real children.
• How models fail, not only how often. Every clip has one known error type, made by a known rule (or none), so missed errors, false alarms and confused categories can be traced to the error type, the rule and the language.
• Does a model really hear it, or just agree? In the not-blind test, models are also asked about errors that are not in the clip. A model that confirms everything it is told is caught; only one that says yes when the error is there and no when it isn't really hears it.
• Languages beyond English. The same categories and mix in every language, including Egyptian Arabic, a dialect with very little child speech data, show whether a model only works well in English.

It is meant for benchmarking and stress-testing models, not for training a clinical tool on its own.

Possible future uses

Tested models could one day help children wait less for support, by pointing families to a speech-language pathologist sooner, and help assess children who grow up with more than one language in each of their languages. This dataset does not show that any model can do this; it is a first step toward testing it.

Dataset structure

One configuration per language:

from datasets import load_dataset
ds = load_dataset("ABHammad/Vokidoo-dataset", "en", split="train")

language	normal	substitution	omission	addition	stuttering	total
Egyptian Arabic (arz)	74	33	36	31	31	205
English (en)	85	34	32	34	34	219
German (de)	83	34	34	34	32	217
Spanish (es)	79	36	36	36	36	223
all	321	137	138	135	133	864

Planned mix in every language: 40% typically developing, 15% each substitution, omission, addition and stuttering. A clip that fails a quality check is left out, so the counts above can differ a little from the plan.

Fields

field	description
audio	16 kHz mono recording
disorder_class	typically_developing or speech_disorder
disorder_type	articulation, phonological, stuttering, or typically_developing
disorder_symptom	substitution, omission, addition, stuttering, or typically_developing
transcription	the words the child means to say
age	the child's age in years (it sets the voice's pitch and size, and the level of the sentences)
gender	the child's gender (the voice used); half girls, half boys within every symptom
clip_id	unique id of the clip
language	language code
target_ipa	correct pronunciation (IPA)
realized_ipa	the pronunciation spoken, with the error (for stuttering the stutter is in the audio only)
process	the rule applied: e.g. fronting, s→θ, cluster_reduction, final_epenthesis, repetition
tts_voice	the synthetic voice
duration_s	length in seconds

Ages (clips per age group)

language	5-6	7-8	9-10	11+
arz	39	54	63	50
de	38	61	64	55
en	38	59	67	56
es	40	58	70	56

Gender (clips)

language	F	M
arz	95	111
de	111	107
en	106	114
es	113	111

Labels

• Articulation error: one sound produced wrongly every time (e.g. a lisp, s → θ).
• Phonological error: a pattern across a class of sounds (e.g. fronting: k, g → t, d); it always changes at least two sounds, so the pattern can be heard.
• Substitution / omission / addition: a sound replaced, left out, or an extra sound (vowel) added.
• Stuttering: one kind per clip, 1-3 times at word starts: a repetition of part of a word or a whole word (b-b-ball, I - I - I want), a prolongation (sssun), or a block (a silent stop before a word, then a hard, pushed start).

How the clips were made

1. Text. Age-appropriate words and short sentences in each language, written by a language model (Anthropic Claude) and filtered: no loanwords, no names; American English, Standard German (Hochdeutsch), Spanish from Spain (Castilian), and Egyptian colloquial Arabic (Cairo), not Modern Standard Arabic.
2. Pronunciation. Each text is converted to IPA (espeak-ng; for Egyptian Arabic, written by the language model with Egyptian pronunciation, checked symbol by symbol and checked twice to be everyday Cairo Arabic).
3. Error. The error is written into the IPA by rules based on typical developmental patterns in each language (e.g. fronting, stopping, gliding, cluster reduction, final-consonant deletion, epenthesis; for Egyptian Arabic also de-emphasis and glottal replacement). A clip labelled as a disorder only gets an error that is no longer typical at the child's age.
4. Speech. English, German and Spanish: a Microsoft Azure neural voice of the language reads the text, and the word(s) with the error are spoken from their IPA. Egyptian Arabic: Google Gemini's voice (Azure has no Egyptian dialect), reading the Arabic text with the word carrying the error respelled in Arabic letters, with no instruction (it recognises the Egyptian words itself); the voice is made child-like (pitch and vocal-tract length set from the child's age), stuttering is added to the audio, and light room sound.

Quality control

A clip is published only if it passes every check:

1. Phoneme recognition (wav2vec2): the intended sounds, and the error, are heard where they should be.
2. Audio check: decodes; 16 kHz mono; 0.5-20 s; not silent; not clipped; sounds like speech, not a tone or noise; no duplicates.

Clips that fail are generated again with other words. None of the benchmarked models was used to select the clips, so their scores below are directly comparable.

Benchmark results

Three models judged every clip blind: they hear the recording (with the intended words and the language) and must say whether the speech is normal or which mistake it has: substitution, omission, addition or stuttering. They were then asked not blind: told what the clip is meant to contain, can they confirm it is there?

model	type
gemini-3.8-flash	Google Gemini, audio-native multimodal model
gemini-4-argon	Google Gemini, audio-native multimodal model
nemotron-asr	NVIDIA Nemotron ASR (speech-to-text; its transcript is compared with the intended words)

Blind: fully right (%), the clip's category named correctly

language	gemini-3.8-flash	gemini-4-argon	nemotron-asr
English	80.0	90.5	50.5
German	72.9	82.1	49.1
Spanish	82.1	96.9	54.5
Egyptian Arabic	73.8	78.6	—
all languages	77.3	87.2	51.4

Blind: disorders caught (%), clips with a mistake flagged as a disorder

language	gemini-3.8-flash	gemini-4-argon	nemotron-asr
English	74.8	96.3	88.9
German	69.6	89.6	89.6
Spanish	77.2	97.2	86.2
Egyptian Arabic	69.7	80.3	—
all languages	72.9	91.0	88.2

Blind: normal speech left alone (%), clean clips judged normal

language	gemini-3.8-flash	gemini-4-argon	nemotron-asr
English	96.5	85.9	69.4
German	88.0	75.9	65.1
Spanish	98.7	97.5	83.5
Egyptian Arabic	97.3	89.2	—
all languages	95.0	86.9	72.5

Not blind: confirms the intended content (%), told what the clip contains

language	gemini-3.8-flash	gemini-4-argon	nemotron-asr
English	94.1	96.4	73.2
German	95.0	93.1	72.5
Spanish	96.0	99.6	76.3
Egyptian Arabic	89.2	80.6	—
all languages	94.3	94.4	74.0

Not blind: rejects what isn't there (%), asked on 20% of the clips about an error that is not in them (or, for a clip with an error, asked as if it were clean); the right answer is no

language	nemotron-asr
English	80.5
German	80.6
Spanish	83.9
Egyptian Arabic	—
all languages	81.5

Not blind: balanced (%), the average of the two; a model that always says yes gets 50%

language	nemotron-asr
English	76.8
German	76.6
Spanish	80.1
Egyptian Arabic	—
all languages	77.8

Across all languages, gemini-4-argon is the most accurate blind judge (87.2% fully right; guessing among the 5 answers would give about 20%). Notes:

• Nemotron writes the word a speaker meant, so it hides many mispronunciations, and stuttering most of all; it has no Arabic, so Egyptian Arabic is compared between the two Gemini models.
• Full tables (per language, per symptom, confusion counts) are in results/.

Intended use

• Benchmarking speech, audio and multimodal models on detecting and naming children's speech-sound errors, across languages.
• Studying how such models fail: missed errors, false alarms, confused categories.

Not for: diagnosis, screening or any clinical decision about a real child, or as the only training data for a clinical tool.

Limitations

• Synthetic voices are not children. Real children's errors are often subtler and less consistent; voice quality, rhythm and accents differ.
• Errors follow rules of typical development. Labels describe what was generated and confirmed by automatic checks; they were not rated by a speech-language pathologist.
• The automatic checks use models that may share blind spots with the models being tested.
• Egyptian Arabic long e and o (and zˤ) are spoken as the nearest sounds the voice has (i, u, z), so they sound slightly less Egyptian; words without them are preferred, and every text is checked twice to be everyday Cairo Arabic.
• Mistakes follow age norms: a clip labelled as a disorder only has a mistake that is no longer typical at the child's age. The age values are drafts from the speech-development literature for each language.
• Stuttering is simulated: one kind per clip, 1-3 times, at word starts. Real stuttering also has physical tension and other behaviours that audio synthesis can't show.
• Egyptian Arabic is spoken by Google Gemini's voice, and the two judges for Egyptian Arabic are also Gemini models: speech from the same model family may be easier for them. Their scores compare fairly with each other, but less directly with the other languages.
• Voices: English and German girls' clips use Azure's child voices; every other clip is an adult voice made younger (pitch and formants), which can sound slightly processed.

Citation

@misc{slp25_synthetic,
  title  = {SLP-25: synthetic child speech with speech-sound errors},
  author = {ABHammad},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ABHammad/Vokidoo-dataset}
}