Is Semi-Automatic Transcription Useful in Corpus Creation? Preliminary Considerations on the KIParla Corpus

Martina Simonotti; Ludovica Pannitto; Eleonora Zucchini; Silvia Ballarè; Caterina Mauri

cs.CLMar 17, 2026

Is Semi-Automatic Transcription Useful in Corpus Creation? Preliminary Considerations on the KIParla Corpus

Authors: Martina Simonotti, Ludovica Pannitto, Eleonora Zucchini, Silvia Ballarè, Caterina Mauri

Organizations: University of Bologna, Bologna - Italy · Masaryk University, Brno - Czech Republic

Abstract

This paper analyses the implementation of Automatic Speech Recognition (ASR) into the transcription workflow of the KIParla corpus, a resource of spoken Italian. Through a two-phase experiment, 11 expert and novice transcribers produced both manual and ASR-assisted transcriptions of identical audio segments across three different types of conversation, which were subsequently analyzed through a combination of statistical modeling, word-level alignment and a series of annotation-based metrics. Results show that ASR-assisted workflows can increase transcription speed but do not consistently improve overall accuracy, with effects depending on multiple factors such as workflow configuration, conversation type and annotator experience. Analyses combining alignment-based metrics, descriptive statistics and statistical modeling provide a systematic framework to monitor transcription behavior across annotators and workflows. Despite limitations, ASR-assisted transcription, potentially supported by task-specific fine-tuning, could be integrated into the KIParla transcription workflow to accelerate corpus creation without compromising transcription quality.

Figures & tables

speaker	start	end	transcription	translation
SP2	3.14	6.07	sì (.) non era pesantissima	yes (.) it wasn’t too rich
SP1	6.34	7.52	per niente [(ha-)]	not at all [(the-)]
SP2	7.22	10.09	[uni]ca cosa, ho esagerato un po’ con l’olio forse però	[the on]ly thing is, I exaggerated a bit with oil maybe but
SP1	10.85	12.28	ma nell’impasto? no fuori,	but in the dough? no outside,
SP2	12.46	15.76	°fuori° (.) nell’im[pasto ce ne ho messo il] giusto però in padella	°outside° (.) in the d[ough I put just] enough but in the pan
SP3	13.39	14.36	[nella padella]	[in the pan]

Table 1: Extract of a transcription of a free conversation from KPS021, KIPasti module ( Mauri et al., 2024a ) . Seven transcription units are shown, with their respective time boundaries, uttered by three different speakers. Text is enriched with Jefferson-style notation.

Symbol	Meaning
word,	Weakly rising intonation
word?	Rising intonation
word.	Falling intonation
wo:rd	Prolonged sound
(.)	Short pause
word=word	Prosodically linked units

Table 2: Transcription conventions

Conversation type	Extract	Tokens/min	Types/min
Free conversation	A	183.23	64.39
	B	199.31	78.08
Semi-structured interview	A	172.30	52.85
	B	170.80	51.04
Interviews with L2 material	A	199.45	61.49
	B	180.61	70.15

Table 3: Tokens and Types per minute in “Gold” transcriptions for the 6 considered 10-minute extracts.

speaker	start	end	transcription	translation
SP1	684.82	687.56	il tuo posto del cuore di bologna	the place of your heart in Bologna
SP1	688.01	689.78	non so anche solo un bar	I don’t know maybe just a cafeteria
SP2	689.80	690.29	[mh]	[mhm]
SP1	689.83	691.95	[a cui se]i particolarmente affezionata,	[that you]’re particularly fond of
SP2	692.63	693.13	mh	mhm
SP2	695.10	697.55	e:h forse: mh:h	e:h maybe uh:m

Table 4: Extract of a transcription of a semi-structured interview from PBA001, ParlaBO module ( Mauri et al., 2024b ) ). Text is enriched with Jefferson-style notation.

speaker	start	end	transcription	translation
SP1	830.44	832.43	parla- m:h che lingua parlate?	you spea- uh:m which language do you speak?
SP2	832.79	835.93	[bangla] sì è la lingua:: bangla che parlan[o]	[bangla] yes it’s the bangla:: language they spe[ak]
SP1	832.79	833.33	[quale li:-]	[which la:-]
SP1	835.86	836.34	[e:]	[a:nd]
SP2	836.17	839.12	[p]erò eh tutti quanti lo sanno anche italiano °tranquilla[mente°]	[b]ut eh everybody has no problem with Italian °as [well]°
SP1	838.72	843.35	[okay] e lo mi- mischiate un po’ [con l’italiano] quando parlate fra di [voi o]	[okay] and you mi- mix it [with Italian] a bit when you speak among each [other or]

Table 5: Extract of a transcription of an L2 interview from SBIB003, StraParlaBO module (see Zucchini et al. accepted ; Mauri et al. 2025 and Zucchini et al. (2026) ). Text is enriched with Jefferson-style notation.

	Manual Phase					ASR-assisted Phase
Transcriber	conversation	30 min	60 min	90 min	120 min	conversation	30 min	60 min	90 min	Final
A	L2-interview 2	60	129	193	232	L2-interview 1	74	110	153	195
E	interview 1	117	268	378	503	free-conversation 1	113	257	360	465
I	L2-interview 1	69	138	229	254	interview 1	71	133	206	290
O	free-conversation 1	120	220	396	406	L2-interview 2	90	180	254	304
L	L2-interview 1	51	118	224	280	interview 1	41	79	130	175

Table 6: Audio extracts assigned to both expert (top pane of the table) and novice transcribers (bottom pane) in the two phases of the experiment, along with transcription times comparison. In the ASR-assisted phase only three measures were collected, as the first part of the session was used to pre-process the Whisper output. Measures are expressed in seconds. See also Figures 3 and 4 in Appendix.

		tokens per minute			TUs per minute
Module	$\Delta$	0	1	2	0	1	2
interview	$\Delta_{M}$	$\downarrow$ 5	$\downarrow$ 8	$\downarrow$ 16	$\uparrow$ 4.67	$\uparrow$ 11.33	$\uparrow$ 20.17
interview	$\Delta_{A}$	$\uparrow$ 2	$\downarrow$ 1	$\downarrow$ 8	$\uparrow$ 7.67	$\uparrow$ 18.33	$\uparrow$ 28.17
free-conversation	$\Delta_{M}$	$\uparrow$ 6	$\uparrow$ 15	$\uparrow$ 29	$\uparrow$ 6.83	$\uparrow$ 11.00	$\uparrow$ 14.33
free-conversation	$\Delta_{A}$	$\uparrow$ 4	$\uparrow$ 22	$\uparrow$ 41	$\uparrow$ 1.83	$\uparrow$ 2.00	$\uparrow$ 0.33
L2-interview	$\Delta_{M}$	$\uparrow$ 2	$\uparrow$ 15	$\uparrow$ 16	$\downarrow$ 7.47	$\downarrow$ 16.67	$\downarrow$ 20.53

Table 7: Deltas in tokens per minute and number of transcription units per minute across modules and transcription methods (Gold vs. Manual/ASR-assisted) at time intervals 0–2 minutes. Positive values are identifies by a

\uparrow

symbol while negative values by a

\downarrow

one.

		Manual		ASR-assisted
Transcriber	Expert	Data	WER	Data	WER	$\text{WER}_{M}-\text{WER}_{A}$
A	Yes	L2-interview	26,35	L2-interview	15,14	11,21
E	Yes	interview	15,92	free-conversation	17,99	-2,07
I	Yes	L2-interview	20,09	interview	24,46	-4,37
L	No	L2-interview	25,04	interview	22,74	2,3
M	No	interview	20,9	free-conversation	29,77	-8,87

Table 8: WER Analysis by Transcriber Expertise, Transcription Method and Deltas

	severe	mild	non-severe
novice_manual	75%	20%	5%
novice_ASR	52%	40%	8%
expert_manual	54%	42%	4%
expert_ASR	60%	40%	0%
reviewer	38%	41%	21%

Table 9: Percentage of overlap errors according to type and expertise

Figure 1: Percentage of overlap errors, novices’ transcriptions

Figure 2: Percentage of overlap errors, experts’ transcriptions

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Participant	Transcription experience
O	1:06:23
E	2:53:32
A	1:00:51
I	0:46:48
R,M	few minutes
L,V,N,U,S	only trial transcriptions

Appendix

Table 10: Transcription experience of the eleven participants of our experiment, expressed in hour:minutes:seconds format.

Figure 3: Transcribed minutes by experts (in blue) and novices (in orange) over time during phases 1 (continuous lines) and 2 (dashed lines)

Figure 4: Transcribed minutes by experts (in blue) and novices (in orange) over time for the three different data sources.

Figure 5: Marginal effects of expertise level (expert/novice), data type (free-conversation/interview/L2-interview), and experimental phase (manual/ASR-assisted) on predicted transcription accuracy

Explore similar work

Easper: An Accessible ASR Pipeline for Language Documentation
Aug 12, 2026Aso Mahmudi, Ting Dang, Ekaterina Vylomova +1ASR Evaluation Automatic Speech Recognition
School of Computing and Information Systems, The University of Melbourne, Australia · School of Languages and Linguistics, The University of Melbourne, Australia
WorldSpeech: A Multilingual Speech Corpus from Around the World
May 9, 2026Antonis Asonitis, Luca A. Lanzendörfer, Frédéric Berdoz +1Automatic Speech Recognition Low-Resource Speech Recognition
ETH Zurich
Transcribing Children's Speech: ASR Performance and Obtaining Reliable Orthographic Transcriptions
Apr 10, 2026Gus Lathouwers, Lingyun Gao, Catia Cucchiarini +1Automatic Speech Recognition ASR Evaluation
Radboud University, the Netherlands

Aug 12, 2026J/K move · Enter open · S save

Easper: An Accessible ASR Pipeline for Language Documentation

Aso Mahmudi, Ting Dang, Ekaterina Vylomova, Nick Thieberger

School of Computing and Information Systems, The University of Melbourne, Australia · School of Languages and Linguistics, The University of Melbourne, Australia

ASR Evaluation Automatic Speech Recognition Low-Resource Speech Recognition

Open paper AskPDF arXiv

Audio transcription is a critical bottleneck in language documentation. While multilingual Automatic Speech Recognition (ASR) models like Whisper offer solutions, field linguists often lack the expertise to utilise them. We present Easper, an open-source, no-code workflow enabling linguists to iteratively fine-tune ASR models via cloud resources directly from ELAN annotations. Deploying ASR also raises a cold start problem: deciding which recordings to transcribe first to bootstrap an accurate model. Using Easper, we evaluate transcription prioritisation strategies on three Vanuatu languages (Bislama, Nafsan, Nguna). We fine-tune models by recording session, comparing Character Error Rate trajectories when prioritising acoustic cleanliness versus linguistic richness. We demonstrate that prioritising lexically rich narratives and increasing acoustic-phonetic repetition, even in noisy environments, leads to faster improvements in transcription quality.

Abstract

Explore similar work

Easper: An Accessible ASR Pipeline for Language Documentation

WorldSpeech: A Multilingual Speech Corpus from Around the World

Transcribing Children's Speech: ASR Performance and Obtaining Reliable Orthographic Transcriptions