Introduction
Automatic speech recognition (ASR), powered by generative artificial intelligence (Gen-AI) and based on large language models (LLM), has had a remarkable impact on professional reporting during the last five years. There has been lively discussion about the role and influence of ASR on the methods and workflow of reporting in different professional contexts (e.g. Haanen et al. 2020; Kawahara et al. 2020; Schelhaas 2021; Lombard 2022; Varisto & Kuronen 2022; Bruno 2023; Eugeni 2023; Granja 2023; Nielsen 2024; Pagano 2024; Smith & Reid 2024; Kerr 2025). However, there has been much less discussion about the effect of ASR on linguistic and editorial practices of reporting. This is the perspective of this article (see also Perreira and Granja, in this issue).
My focus is on the Records Office of the Finnish Parliament, which has used ASR to make first drafts for parliamentary reporters since 2019. We use an ASR software made by a Finnish company called Lingsoft. The software has been trained with our text and video materials, partially modified for our needs and integrated in our audio recording system. Our office has named it Pulina, or ‘Babble’. Since spring 2025, Pulina has had a new, improved language model that has been much more accurate than the previous version. As a new addition, it has also started to make some independent editing suggestions, based on, for example, earlier Finnish parliamentary reports in its training material. For this article, I have used the observations and discussions about Pulina in our office to analyse how the AI-based ASR automatically edits the language of the draft report.
From punctuation to phonological and morphological standardisation
The ASR software in the Finnish Parliament is mainly a transcription tool that turns the MPs’ speeches into written text. However, it also makes various editorial suggestions. First, it introduces punctuation to the draft report. It adds, for example, periods in perceived sentence breaks, exclamation points in the ritual forms of address (Arvoisa puhemies! ‘Honourable Chairman!’) and question marks in questions. It attempts to add commas according to Finnish orthography, and occasionally even adds colons in introductory sentences (see example 1).
Example 1.
lopuksi aivan lyhyesti muista lisätalousarvioehdotuksen yksityiskohdista:
‘finally, about other details of the supplementary budget proposal:’
Pulina does make frequent mistakes in punctuation. However, the reporters have found it useful, because they feel that it is easier to correct the mistakes than add all punctuation manually in the report.
Besides orthography, AI-based ASR in the Finnish Parliament also does some phonological and morphological standardisation according to the norms of the Finnish standard language. In example 2, it has filled in the plural form (-vat) in the 3rd person verb form, usually absent in everyday spoken Finnish (in bold).
Example 2.
pysyvään oleskelulupaan voisivat oikeuttaa ainoastaan korkeat tulot
‘only high incomes could qualify for a permanent residence permit’
This addition is also made by the parliamentary reporters, because they see the standard form as more prominent in written texts. However, the reporters do not follow the standard language when they think that it is unnecessary or that it excessively changes the rhetoric or style of the original speech. In such cases, Pulina might suggest standardisation against the office guidelines (e.g. me mennään à me menemme ‘we go’). Also, the AI does not standardise the report systematically. These cases show that it is necessary to have an alert human reporter to review AI’s editorial suggestions according to the office guidelines. (On the editorial principles of the Finnish parliamentary report, see Voutilainen 2023).
Transforming sentence structures
Pulina also occasionally standardises spoken language syntax to fit the norms of Finnish standard language. In example 3, it has added two small words that do not carry significant meaning (on ‘is’ and se ‘it’) in order to achieve a standard sentence structure.
Example 3.
sinällään se, mikä on hyvää Suomen näkökulmasta, on se, että – –
‘basically what is good from Finland’s point of view is that – -’
This type of change is often made by the reporters as well, because they see that it increases readability without substantially changing the meaning or style of the speech. Small changes in word order are also common to retain the presumed intended meaning in transcribed speech (e.g. eduskunta myös pidetään tiiviisti mukana ‘the parliament is also kept tightly along’ à myös Eduskunta pidetään tiiviisti mukana ‘also the parliament is kept tightly along’). Sometimes these changes are useful, but they are occasionally too extensive for the office guidelines.
While transforming spoken syntax to fit the conventions of the written standard language, the AI-based ASR software also fades out the spokenness of the reported speeches. For example, it often erases word-searches, stuttering and real-time planning from the speaker. In example 4, the speaker changes the sentence structure at the beginning of the sentence. Pulina unifies the sentence structure to fit the ending, which is the MP’s final choice.
Example 4.
ehkä kysymykset oli vähän sitten… vei minut harhaan → ehkä kysymykset vähän sitten veivät minut harhaan
‘perhaps the questions, then, were a little… led me astray’ → ‘perhaps the questions, then, led me a little astray’
Pulina also removes some self-repairs by the speaker: it removes the wrong word and leaves in the correction that the speaker adds afterwards (e.g. toimeentulotukeen… toimeentuloon ‘income support… income’ → toimeentuloon ‘income’).
The editorial choices in examples 3 and 4 follow the office guidelines. However, this is not always the case. For example, the AI frequently mistakes lists or consequent, structurally similar clauses for self-repair (e.g. projekteja, hankintoja ‘projects, acquisitions’ → hankintoja ‘acquisitions’) and might even remove several meaningful words because of it (e.g. heikentää rahoituksen pienentymisen myötä palvelujen saatavuutta → heikentää palvelujen saatavuutta ‘weakens the availability of services alongside diminishing funding’ → ‘weakens the availability of services’). It also occasionally removes intentional and rhetorically relevant repetition for the same reason.
From over-editing to hallucinations
AI-based ASR often also suggests edits that are not justified according to the office guidelines. In example 5, Pulina has removed a completely correct formulation tutkimusryhmineen ‘with his research group’ and changed it to ja tutkimusryhmä ‘and his research group’.
Example 5.
emeritusprofessori M. O. tutkimusryhmineen esitteli → emeritusprofessori M. O. ja tutkimusryhmä esittelivät
‘professor emeritus M.O. with his research group presented’ → ‘professor emeritus M.O. and the research group presented’
The replacement expression by Pulina is more common but completely unnecessary. It also changes the meaning of the expression and gives a different feel to how active and independent the research group has been according to the MP. In an even bigger case of over-editing, Pulina changes an uncommon but grammatically correct verb form antanee ‘will probably give’ into a more common antanut ‘has given’, which has a completely different meaning and is incorrect in the context.
Sometimes AI adds extra words or ‘hallucinates’, as it is often called. In example 6, Pulina has added a completely new word samalla ‘at the same time’ which was never said in the speech.
Example 6.
sillä voidaan samalla säästää jopa 20 prosenttia työajasta
‘with it we can save as much as 20 percent of the time of labour at the same time’
Over-editing and hallucinations by the AI-based ASR highlight the fact that the human reporter must always carefully inspect the draft report.
Experiences of effect and control
The ASR based on Gen-AI and LLM and used in the Finnish Parliament is very accurate and efficient. It reduces the manual labour of typing and allows the reporters to put more emphasis on editorial decisions. However, its language model leads it to both under- and over-edit the reported speeches. Editing drafts by the ASR software changes editors’ focus from spotting typing errors to examining whole sentences and paragraphs in light of the presumably intended meaning of the speaker and on the office guidelines. Many editors feel that by foregrounding some editorial decisions before others, the ASR might steer their decision-making process, because it is easy to just keep the AI’s choice when other options do not necessarily come to mind. Other colleagues feel that the ASR is merely a useful technological tool and does not affect their decision-making.
I have also observed that the AI’s training material might make it linguistically conservative. By having a great amount of old reports to learn from, it also echoes the earlier editorial solutions of recent and past reporter generations. In practice, this might mean that new editorial principles meet some resistance by the AI. This requires the reporters to be conscious of the impact that the AI-based ASR software tools may have on parliamentary reporting.
Eero Voutilainen is a parliamentary reporter who leads the linguistic accessibility team at the Parliament of Finland. He is also Tiro’s editor-in-chief. He wants to acknowledge the invaluable input of his colleagues who gave their insight and examples about the themes of this article: Elisa Aaltonen, Annamari Koskelo, Riikka Kuronen, Sanna Niemi, Kalle Niemimaa, Vuokko Ranki, Olli Wainikka, and other staff of the Records Office of the Finnish Parliament.
References
Bruno, M. (2023). Respeaking Skills: A Comprehensive Approach. – Tiro 2/2023: https://tiro.intersteno.org/2023/12/respeaking-skills-a-comprehensive-approach/
Eugeni, C. (2023). Mice, Machines and Men. – Tiro 1/2023: https://tiro.intersteno.org/2023/07/mice-machines-and-men/
Granja, P. (2023). The Portuguese Parliamentary Reporters’ Experience with Automatic Speech Recognition Systems. – Tiro 2/2023: https://tiro.intersteno.org/2023/12/the-portuguese-parliamentary-reporters-experience-with-automatic-speech-recognition-systems/
Haanen, M., S. Hoogzand & M. Petrina-Bosch (2020). Live subtitling at the Dutch House of Representatives. – Tiro 1/2020: https://tiro.intersteno.org/2020/05/live-subtitling-at-the-dutch-house-of-representatives/
Kawahara, T., S. Ueno & M. Morikawa (2020). Transcription System using Automatic Speech Recognition in the Japanese Parliament. – Tiro 1/2020:https://tiro.intersteno.org/2020/05/transcription-system-using-automatic-speech-recognition-in-the-japanese-parliament/
Kerr, D. (2025). Harnessing Whisper at the Legislative Assembly of British Columbia: A User-Driven Approach to AI-Supported Parliamentary Reporting. – Tiro 1/2025: https://tiro.intersteno.org/2025/06/harnessing-whisper-at-the-legislative-assembly-of-british-columbia-a-user-driven-approach-to-ai-supported-parliamentary-reporting/
Lombard, C. (2022). Experimenting with Automatic Speech Recognition in the Houses of the Oireachtas (Irish Parliament). – Tiro 1/2022: https://tiro.intersteno.org/2022/07/experimenting-with-automatic-speech-recognition-in-the-houses-of-the-oireachtas-irish-parliament/
Busch Nielsen, L. (2024). Respeaking in the Danish Parliament’s Hansard. – Tiro 1/2024: https://tiro.intersteno.org/2024/06/respeaking-in-the-danish-parliaments-hansard/
Pagano, A. (2024). Verbatim vs. Edited Live Parliamentary Subtitles. – Tiro 2/2024: https://tiro.intersteno.org/2024/12/verbatim-vs-edited-live-parliamentary-subtitles/
Schelhaas, D. (2021). Cloud-based ASR solutions: Some notes for professional reporters. – Tiro 2/2021: https://tiro.intersteno.org/2021/12/cloud-based-asr-solutions-some-notes-for-professional-reporters/
Smith, C. & K. Reid (2024). Towards a Substantially Verbatim Official Report Using Automatic Speech Recognition. – Tiro 1/2024: https://tiro.intersteno.org/2024/06/towards-a-substantially-verbatim-official-report-using-automatic-speech-recognition/
Varisto, N. & R. Kuronen (2022). A Good Servant but a Bad Master. Introducing ASR at the Parliament of Finland. – Tiro 2/2022: https://tiro.intersteno.org/2022/12/a-good-servant-but-a-bad-master-introducing-asr-at-the-parliament-of-finland/
Voutilainen, E. (2023). Written representation of spoken interaction in the official parliamentary transcripts of the Finnish Parliament. – Frontiers in Communication 8/2023: https://www.frontiersin.org/journals/communication/articles/10.3389/fcomm.2023.1047799/full
