Introduction
This article examines the experience of implementing STAAR (Sistema de Transcrição Automática da Assembleia da República), focusing on an unexpected challenge that emerges from artificial intelligence-based automatic speech recognition (ASR) technology: the occasional misinterpretation of speech, from minor misrecognitions to the invention of entirely new words. It should be noted that the humour in some of the examples presented in the article may be lost in translation. Since the original slip-ups occurred in Portuguese, I have provided English translations. However, since these mistakes stem from mishearing or mispronunciation, a few nuances will be inevitably lost along the way for non-Portuguese readers.
So what is STAAR? STAAR, the ASR technology used in the Portuguese Parliament (Assembleia da República, referred as AR from now on), was designed for AR and based on the Whisper model (Nascimento et al., 2024; Nascimento, 2023). It is currently used in Divisão de Redação (AR’s reporting division) for the official journal, the Diário da Assembleia da República – I Série, in the transcription of plenary sessions, as well as other parliamentary meetings.
STAAR’s secrets
Among STAAR’s very interesting contributions is the suggestion of a secret government. While none of these entities have been confirmed by any official source, we can always imagine what our countries would be like if such mysterious ministries existed.
Therefore, we may have a “second-minister” (“segundo-ministro”, instead of “Sr. Primeiro-Ministro”, or “Mr Prime Minister”), a “Secretary of State and Left Affairs” (“Sr. Secretário de Estado e Assuntos Esquerdos”, instead of “Secretário de Estado e Assuntos Internos”, or “Secretary of State and Internal Administration”), a “Secretary of State for Regal Affairs” (“Sr. Secretário de Estado dos Assuntos Regais”, instead of “Sr. Secretário de Estado dos Assuntos Fiscais”, or “Secretary of State for Financial Affairs”), a “Secretary of State for News”(“Sr. Secretário de Estado das Notícias”, againinstead of “Sr. Secretário de Estado dos Assuntos Fiscais”, or “Secretary of State for Financial Affairs”), a “Minister of Whistling” (“Sr. Ministro do Assobio”, instead of “Sr. Ministro, faça favor”, or “You have the floor, Minister”), and a “Mr Army of Health” (“Sr. Exército da Saúde”, instead of “Sr. Ministro da Saúde”, or “Mr Minister of Health”).
Formal parliamentary speech
STAAR sometimes mixes up expressions from the Regimento da Assembleia da República (the rules of procedure in the Portuguese Parliament), and its main victim is the clarification request, “pedido de esclarecimento”. Therefore, we may have “growth requests” (“pedidos de crescimento”), “requests with distance” (“pedidos com distanciamento”), “ignorance requests” (“pedido de desconhecimento”), and “disclassification requests” (“pedido de desclassificamento”). We may also have a “round of birth politicians”(“Ronda de Políticos do Nascimento”).
At another moment, STAAR misinterprets an interpellation to the board (“interpelação à Mesa”, when a Member of Parliament addresses the President directly to clarify an issue or to distribute a document) as the sentence: “Did I jump on your table?” (“Pulei pela sua mesa?”).
Still in the formalities of parliamentary language, a frequently used phrase by the President when inviting a MP (Member of Parliament) to speak is “faça favor”, corresponding to “if you please”. However, STAAR often misunderstands this polite expression, transforming it into unintentionally comical commands directed at MPs. So an MP may be directed to “do him/herself” (“Faça-se, Sr. Deputado”), to “close the air” (“Muito bem, fecha o ar, Sr.ª Deputada”), to “make jam” (“Faz doce, Sr.ª Deputada”), or to “do it now” (“Faça agora”). The MP might receive a warning that “it is going to rain” (“Sr. Deputado, vai chover”) or that he only has “two minutes for the charade” (“2 minutos para a charada”), and the President might also indicate that he is “fascinated” (“É por estar fascinado”).
Surreal and paranormal moments
There are moments when the machine simply does not recognise a word or an expression, so the result is just near enough. This results in both comical and surreal moments, or, if we are audacious enough, in suggestions of an alternative reality.
For instance, we may have a “National Railway Palace” (when STAAR confuses “paço” and “passe”, which means “pass”); the opposition may accuse the Government of doing a “gymkhana with the octopus” (from the confusion between “polvo” and “povo”, which means “people”); and unconstitutionality may “stink” (from a mix-up between “fede” and “fere”, or “to hurt”).
Petitioners (“peticionários”) are often transcribed as “dictionaries”, and the Parliament galleries can be occupied by a “delegation of dictionaries”. A political party may be in a “state of medication” (rather than a “state of denial”, from STAAR’s confusion between “negação” and “medicação”) and a MP may defend “the abolition of women with a PhD” (“a abolição das doutoradas”, instead of “touradas”, or “bullfighting”).
Finally, STAAR sometimes jumps to paranormal conclusions, referring to “celestial fires” (confusing “celestiais” and “florestais”, or “wildfires”); Portuguese teachers may be abducted by “terrestrial angels” (“anjos terrestres”, instead of “extraterrestres”, or “aliens”); and we may face the “railway evil” (“mal” instead of “modo”, or “mode”).
Rethinking human work with automatic transcription
Highlighting STAAR’s slip-ups does serve a purpose other than to amuse. While artificial intelligence tools are unquestionably useful, these examples reveal the unpredictability of machine work and emphasise the need for human supervision.
When STAAR was introduced, Divisão de Redação had to adapt its workflow. Previously, the first revision stage was performed by reporters who transcribed audio and corrected issues like subject-verb agreement, repetitions, and grammar as they typed the text. This initial stage combined transcription with preliminary editing.
A second review was then conducted to catch any remaining errors and, finally, a team member performed a last read-through without audio reference, known as “lapsos”.
With STAAR, the first stage of revision shifted from typing to revising the transcription from ASR. The second stage of revision is still performed by a proofreader and the final stage is done by an editor.
Although STAAR corrects grammar and eliminates some repetition while maintaining a low word error rate (2%), it still requires a careful revision effort from the members of Divisão de Redação. Automatic transcription has made the revision process more demanding, requiring greater attention to errors, especially because typos are easier to overlook in prewritten text.
Conclusion
While there are no major gains in overall time savings, the technology does enable the processing of larger volumes of text, with the quality of transcription depending on speaker clarity and the complexity of the subject matter. Another major advantage from the adoption of ASR technology is the decrease in physical strain due to reducing manual typing.
Workflows now highlight revision, with more staff being trained to do the second level of revision (proofreading) and many members of the team rotating through different revision stages.
Despite all its advantages, AI is far from being able to work alone. Effective teamwork comes from collaboration, not just between people but between people and machines. While AI can process data at extraordinary speed, only humans can provide the judgment and critical thinking needed to complete the task.
Ana Luísa Reis is a reporter in Divisão de Redação in the Portuguese Parliament, and this article would not have existed without the contributions of Ana Rita Pereira, Inês Fernandes, Inês Vila, Manuel Moreno, Paulo Granja, Susana Pinto, and Teresa Silva Carvalho.
Note
The article’s examples of mix-ups have been taken from automatic transcription files for several issues of the Diário da Assembleia da República – I Série, as well as transcriptions from committee meetings.
References
Nascimento, P. M. D. (2023). Sistema de Transcrição Automática de Debates Parlamentares no Contexto da Assembleia da República Portuguesa. [MSc dissertation, ISCTE-IUL]. Repositório ISCTE-IUL. https://repositorio.iscte-iul.pt/bitstream/10071/31035/1/master_pedro_delgado_nascimento.pdf
Nascimento, P., J. C. Ferreira, & F. Batista (2024). Automatic transcription system for parliamentary debates in the context of Assembly of the Republic of Portugal. International Journal of Speech Technology, 27, 613-635. https://doi.org/10.1007/s10772-024-10126-4.
