Published Tuesday 19 August 2025
The National Film and Sound Archive of Australia is custodian of the national audiovisual collection – a dynamic and living record of Australia’s cultural, social and creative life, which continues to evolve alongside each major technological shift in audiovisual creation and experience.
Spanning the 1890s to the present, our collection is vast. There are more than 600,000 works in the NFSA collection featuring Australian speech. These include films, TV and radio shows, podcasts, oral histories, news and current affairs and more; everything from vintage television commercials from the 1960s to community radio broadcasts from last week. It would take more than 20 years for one person to listen to it all.
As the volume and complexity of audiovisual content continues to grow, the NFSA is innovating behind the scenes, investing in infrastructure and developing tools to support our future as a digital-led archive.
Conceived and developed by the NFSA, Bowerbird is a machine learning-enabled mass audio and video transcription engine. By understanding, transcribing and documenting hundreds of thousands of hours of collection material, Bowerbird enables staff, researchers and media industry professionals to discover and understand what is in the collection, and to uncover multitudes of previously hidden stories.
With such a large collection of Australian speech – complete with references to flaming galahs, bludgers, dunnies, bogans, avos, arvos and the ABC, and to towns like Woolloomooloo and Warrnambool – the NFSA was not confident in the value of investing in an American-trained engine. It was clear that we needed to develop our own tools, trained on Australian speech and ready to sample our immense digitised collection. To support this objective, Bowerbird is being trained to understand Australian-accented English more accurately and to identify the quintessentially Australian places, people and things that other technologies miss.
Challenges
The NFSA faced significant challenges in making its enormous collection of audio and audiovisual content searchable and accessible. The main issues included:
- Inaccessibility: The collection contains linear content (audio and audiovisual) that was not transcribed or fully documented, making it difficult to find relevant information.
- Inaccuracy: Existing AI models generating transcription (speech-to-text) are not tailored to Australian English or Australian content.
- Resource-Intensiveness: Manual transcription is time-consuming and resource intensive. It would cost almost $10 million to manually transcribe all collection material.
These challenges impacted the NFSA's ability to fully open up the national collection, hindering research, industry reuse and public engagement.
In 2023, the NFSA developed a series of internal AI pilot projects which included technologies such as machine transcription, computer vision, retrial augmented generation and graph databases. These proof-of-concept prototypes were evaluated and the most successful, including Bowerbird, were developed into enterprise applications in 2024 and 2025.
With the development of Bowerbird and new web search tools, the NFSA was able to transcribe, document and link more than 150,000 hours of content, representing millions of unique stories, for a fraction of the cost of commercial human or automated transcription.
Further, the development of Bowerbird needed to adhere to our principles for machine learning and artificial intelligence creation and use which established three strategies to guide the NFSA: maintain trust, build effectively and transparently, and create public value.
Today, Bowerbird is in daily use at the NFSA and is also employed by a select group of Australian cultural partners to transcribe audio and video from their own collections.
Next Steps
A fully searchable collection of millions of audiovisual works has created a significant positive impact for the NFSA as well as external researchers, academics and media industry users. Especially as new digital formats are increasingly networked and data-heavy, requiring specialised systems to preserve them for the future.
The NFSA is creating a database representation of all collection catalogue data, Bowerbird transcriptions and inferred entities, such as places, people and things. This will allow users to find anything in the collection about Skippy, Vegemite, Kylie Minogue or Wollongong, and can recommend surprising connections using a system the NFSA is developing called Graph Assisted Find.
Working with local technology partners, the NFSA will add optical character recognition and scene description to Bowerbird in 2026 to enrich collection metadata and enable new discoveries from within the collection for staff and media industry clients, making one of Australia’s hardest-working cultural collections even more valuable.
