Worked example 8: English story (60 s)

Recording AA_df7b47e1-… · grade 5 · age 10 · enumerator #54 · window 989–1054 s of the session, 60-s clock from the child's first word at 989.1 s

How to read this page. Section 1 is the printed text the child had in front of them, numbered in reading order. Section 2 is the 60-second clip, timed from the child's first word. Section 3 is what the speech-to-text service typed from that clip; the microphone hears the enumerator too, so prompts like "shabash" can appear between the child's words. Section 4 has one row per scorer with a colour chip for every printed word: the enumerator row is the live tablet marking, the official score; the baseline row is the first automatic method, which matches transcript words to printed words by spelling; each model row is that model's own judgement from the audio or from the transcript. Red where the enumerator has green means the scorer flagged too much; green where the enumerator has red means it waved a mistake through; grey means skipped or never reached; faded means no verdict. "Attempted" is the furthest word reached. Section 5's "difference from enumerator" is each scorer's correct count minus the enumerator's: near zero is good, a large positive number is too lenient, a large negative number too harsh. Section 6 is the second human's re-listen.

1. What the child was supposed to read

1. imran 2. woke 3. up 4. early 5. for 6. school 7. today 8. his 9. class 10. was 11. planting 12. trees 13. he 14. took 15. a 16. small 17. plant 18. in 19. a 20. bag 21. the 22. teacher 23. asked 24. imran 25. to 26. show 27. the 28. class 29. how 30. to 31. plant 32. it 33. imran 34. dug 35. a 36. hole 37. put 38. the 39. plant 40. in 41. the 42. soil 43. and 44. put 45. water 46. in 47. it 48. the 49. teacher 50. was 51. happy 52. with 53. his 54. work 55. his 56. friends 57. clapped 58. for 59. him

The study export holds only right/wrong flags, not the text itself, so this text was reconstructed from 30 children's reads with an LLM; a word or two of it may itself be wrong, which would show up as the same "mistake" in every scorer's row.

2. The child's audio

Play the 60-second clip (Drive, Taleemabad accounts)
8 kHz classroom recording from the May 2026 study, upsampled. Folder: worked-examples clips

3. What the speech-to-text service heard

woke up early for school today last class was plant study the teacher read to a small plant in a bag the teacher asked imran to show the class how to plant it imran burst then a hole put the plant in the soil and put water in it slowly the teacher was happy

whisper-1 forced to English · 54 words inside the 60-s window. The enumerator's own words (prompts, "shabash") are included when they fall inside the window.

4. Every scorer's verdict on every word

read correctly wrong skipped or not reached no verdict

Enumerator (tablet, live)

imranwokeupearlyforschooltodayhisclasswasplantingtreeshetookasmallplantinabagtheteacheraskedimrantoshowtheclasshowtoplantitimrandugaholeputtheplantinthesoilandputwaterinittheteacherwashappywithhisworkhisfriendsclappedforhim

53 correct of 57 attempted in 60 s (words after the 60-second point are grey: the export records 0 for both "correct" and "never reached", so they are only known up to the attempted count)

Baseline: speech-to-text + spelling alignment

44 correct of 51 attempted; 4 extra words the child said are ignored

Gemini 3.1 Pro · audio

41 correct of 52 attempted · model note: The student mispronounced several words (e.g., 'woke' as 'walk', 'early' as 'erry', 'put' as 'pat') and the audio cuts off before the end of…

Gemini 3.8 Flash · audio

50 correct of 52 attempted · model note: The child read accurately and steadily up to word 52 before time ran out, only misreading 'trees' and 'dug'.

Gemini 3.5 Flash-Lite · audio

46 correct of 52 attempted · model note: Child read up to 'it' near the end of the text within the 60 seconds.

Gemini 2.5 Flash · audio

59 correct of 59 attempted · model note: The child read all words correctly with clear pronunciation.

GPT audio mini · audio

0 correct of 59 attempted · model note: All words were read incorrectly.

MiMo 2.5 · audio

59 correct of 59 attempted · model note: Child read all words correctly.

Claude Sonnet 5 · transcript

44 correct of 51 attempted · model note: Child skipped initial 'imran' and misread several words (his->last, planting/trees->plant study, he took->the teacher read to, dug->burst); …

GPT-5 mini · transcript

44 correct of 51 attempted · model note: Judgments based on the provided speech-to-text transcript; extra/missing words treated as wrong or skipped accordingly.

Gemini 3 Flash · transcript

43 correct of 51 attempted · model note: The child read up to word 51 with several substitutions and skips, then stopped.

DeepSeek 3.2 · transcript

44 correct of 51 attempted · model note: Child skipped initial 'imran', misread 'his' as 'last', 'planting' as 'plant', 'took' as 'read', 'dug' as 'burst', and skipped final words a…

5. The counts

Scorer	Words correct	Words attempted (furthest word reached)	Difference from enumerator (correct)	Difference (attempted)
Enumerator	53	57	+0	+0
Baseline: speech-to-text + alignment	44	51	-9	-6
Gemini 3.1 Pro · audio	41	52	-12	-5
Gemini 3.8 Flash · audio	50	52	-3	-5
Gemini 3.5 Flash-Lite · audio	46	52	-7	-5
Gemini 2.5 Flash · audio	59	59	+6	+2
GPT audio mini · audio	0	59	-53	+2
MiMo 2.5 · audio	59	59	+6	+2
Claude Sonnet 5 · transcript	44	51	-9	-6
GPT-5 mini · transcript	44	51	-9	-6
Gemini 3 Flash · transcript	43	51	-10	-6
DeepSeek 3.2 · transcript	44	51	-9	-6

6. The second human

The second human (gulrukh) re-listened: on the correct count the reviewer agreed with the enumerator; on the attempted count the reviewer agreed with the enumerator; the enumerator followed the protocol.