Periodic Reporting for period 1 - MIXTAPE (The Podcast Search Engine)
Okres sprawozdawczy: 2019-11-15 do 2020-11-14
Podcasts have seen in the recent years a steady increase in numbers, both in terms of listeners and content providers. With over 42 million episodes available in January 2021, finding the podcast episodes to listen to is one of the main issues that listeners face. The problem is that podcast channels are generally organised in lists and catalogues for listeners to subscribe to, or they are sorted based on popularity. Indeed, in spite of their increasing popularity there are currently no effective ways of finding podcast episodes based on specific interests (i.e. keywords or more complex queries).
In MIXTAPE we improve the way podcasts are searched by allowing listeners to find content based on topics of interest. For the first time it is possible to find individual podcast episodes based on listener preferences and precise interests. To enable this we use speech processing and natural language understanding and add these key components to our existing content management platform. Our ambition is to become the reference search engine for podcasts (i.e the “Google” for podcasts).
In order to achieve this we hired an Innovation Associate (IA) with background machine learning and data analysis techniques to develop further our content analysis capabilities and include speech processing and natural language understanding in our content platform with a dedicated application use case around search of podcast content.
The specific objectives we wanted to achieve by recruiting an IA are to:
(a) Use state of the art techniques in machine learning and data analysis for speech processing and natural language understanding in order to enrich the metadata of podcasts;
(b) Use the additional metadata in order to develop a user friendly search engine for podcasts
In MIXTAPE we improve the way podcasts are searched by allowing listeners to find content based on topics of interest. For the first time it is possible to find individual podcast episodes based on listener preferences and precise interests. To enable this we use speech processing and natural language understanding and add these key components to our existing content management platform. Our ambition is to become the reference search engine for podcasts (i.e the “Google” for podcasts).
In order to achieve this we hired an Innovation Associate (IA) with background machine learning and data analysis techniques to develop further our content analysis capabilities and include speech processing and natural language understanding in our content platform with a dedicated application use case around search of podcast content.
The specific objectives we wanted to achieve by recruiting an IA are to:
(a) Use state of the art techniques in machine learning and data analysis for speech processing and natural language understanding in order to enrich the metadata of podcasts;
(b) Use the additional metadata in order to develop a user friendly search engine for podcasts
The work started with the design of a flexible system architecture. This was gradually implemented by the IA and already after the first 6 months a first end-to-end content workflow could be tested.
Afterwards the focus of the work shifted towards the improvement of the results achieved when analysing the podcast data. For this, several automatic speech recognition (ASR) technologies have been tested and benchmarked against the gold standard available commercially. This helped us select the best solution and which was then further adapted to the needs of the MIXTAPE use cases.
Following this, the platform, with its content intake pipeline has been deployed in a distributed environment, and several large content intake jobs have been triggered in order to identify common issues and bottlenecks. Several issues came out and have been fixed: the platform now is stable and consistent. Any podcast feed can be added to the scheduler to inject the data into the platform and can be updated through time by defining how frequently it has to be updated. In the final months of the project the IA has focused on developing a friendly and modern user interface for accessing the core search functionality of MIXTAPE.
Moreover, throughout the project the IA has successfully completed a number of in-house tailored training activities and was able to seamlessly become part of the company’s team. The IA and the supervisor actively participated in the core training sessions provide by EASME through Ernst and Young, focusing mainly on the innovation aspects of the project. Both the innovation associate and the supervisor passed the final examination and were awarded the innovation management certificate.
The project was disseminated and communicated through various actions. The project was also featured on the IN2 webpage (https://in-two.com/projects/mixtape(odnośnik otworzy się w nowym oknie)). MIXTAPE was also featured on IN2’s yearly newsletter which has a reach of almost 7000 business and research contacts. A flyer and presentation of MIXTAPE were prepared and we participated and disseminated the project during 3 online events.
The main outcome of the project is a software system that is able to ingest podcast feeds, analyse them using state-of-the-art Machine Learning techniques, automatically extracting a full transcript and using this to augment the existing metadata of the podcast and index the content. Based on this improved data and index it is possible to offer an advanced search for podcast episodes based on topics of interest and keywords which goes beyond the existing capabilities offered by major podcast platforms like iTunes and Spotify. IN2 will pursue the market introduction of MIXTAPE as a new freemium SaaS service for consumers. The MIXTAPE search engine is available at: https://rmxtape.xyz/(odnośnik otworzy się w nowym oknie)
Afterwards the focus of the work shifted towards the improvement of the results achieved when analysing the podcast data. For this, several automatic speech recognition (ASR) technologies have been tested and benchmarked against the gold standard available commercially. This helped us select the best solution and which was then further adapted to the needs of the MIXTAPE use cases.
Following this, the platform, with its content intake pipeline has been deployed in a distributed environment, and several large content intake jobs have been triggered in order to identify common issues and bottlenecks. Several issues came out and have been fixed: the platform now is stable and consistent. Any podcast feed can be added to the scheduler to inject the data into the platform and can be updated through time by defining how frequently it has to be updated. In the final months of the project the IA has focused on developing a friendly and modern user interface for accessing the core search functionality of MIXTAPE.
Moreover, throughout the project the IA has successfully completed a number of in-house tailored training activities and was able to seamlessly become part of the company’s team. The IA and the supervisor actively participated in the core training sessions provide by EASME through Ernst and Young, focusing mainly on the innovation aspects of the project. Both the innovation associate and the supervisor passed the final examination and were awarded the innovation management certificate.
The project was disseminated and communicated through various actions. The project was also featured on the IN2 webpage (https://in-two.com/projects/mixtape(odnośnik otworzy się w nowym oknie)). MIXTAPE was also featured on IN2’s yearly newsletter which has a reach of almost 7000 business and research contacts. A flyer and presentation of MIXTAPE were prepared and we participated and disseminated the project during 3 online events.
The main outcome of the project is a software system that is able to ingest podcast feeds, analyse them using state-of-the-art Machine Learning techniques, automatically extracting a full transcript and using this to augment the existing metadata of the podcast and index the content. Based on this improved data and index it is possible to offer an advanced search for podcast episodes based on topics of interest and keywords which goes beyond the existing capabilities offered by major podcast platforms like iTunes and Spotify. IN2 will pursue the market introduction of MIXTAPE as a new freemium SaaS service for consumers. The MIXTAPE search engine is available at: https://rmxtape.xyz/(odnośnik otworzy się w nowym oknie)
In MIXTAPE, the IA has extended the company’s content analysis capabilities and was able to include speech processing and natural language understanding as part of our core content management and publishing platform.
As part of the project, we also benchmarked the MIXTAPE developments using many types of podcast data (e.g. podcasts recorded in a studio, podcasts with background noises, etc) and different ASR technologies. This analysis shows that we are able to get comparable results to commercial speech recognition engines, especially for podcasts of high production quality.
The new software modules for efficient analysis of recorded speech developed by the IA has been a key radical innovation that our company has benefited from as they already become needed by other projects and products that we implement at this time (namely an automatic video chaptering solution as part of the European Data Incubator). In this way MIXTAPE gave us the opportunity to expand our commercial offering into new applications and services.
On the other hand the IA has benefited from comprehensive training both on technical issues as well as innovation and business aspects and as such is now better equipped to advance professionally, having a better position on the labour market.
As part of the project, we also benchmarked the MIXTAPE developments using many types of podcast data (e.g. podcasts recorded in a studio, podcasts with background noises, etc) and different ASR technologies. This analysis shows that we are able to get comparable results to commercial speech recognition engines, especially for podcasts of high production quality.
The new software modules for efficient analysis of recorded speech developed by the IA has been a key radical innovation that our company has benefited from as they already become needed by other projects and products that we implement at this time (namely an automatic video chaptering solution as part of the European Data Incubator). In this way MIXTAPE gave us the opportunity to expand our commercial offering into new applications and services.
On the other hand the IA has benefited from comprehensive training both on technical issues as well as innovation and business aspects and as such is now better equipped to advance professionally, having a better position on the labour market.