Manuel Rivera Morales’ Voice Re-created by AI for the Olympic Games

Manuel Rivera Morales gained nationwide fame in Puerto Rico during the 1960-80s when he covered BSN basketball games live, first on the radio and later on television.

Puerto Ricans have long associated the voice of Rivera Morales with the notable success of their basketball team in the 70s and 80s. Even today, Rivera Morales is considered by many to be the best sportscaster in Puerto Rican history and one of the best in the world.

The DDB agency and Respeecher partnered to revive his memorable voice with the help of artificial intelligence (AI). The project was devoted to the National Women's Basketball Team, which made its debut at the Olympic Games for the first time, playing against China.

The idea was for the sportscaster to broadcast the match, which would be an absolute surprise for the nation. 

The result was an emotional performance in which the voice of Rivera Morales narrated the Puerto Rico match against China on Tuesday, July 27, sharing the broadcast with the commentator Natalia Meléndez.

This project has been recently recognized by the SME Awards 2022 and has won three awards in the following categories:

1. For Innovative Use of Technology - gold award
2. Social media - bronze award
3. Branded content - gold award

SME is an association recognizing the effective integration of social media in marketing campaigns, public relations, advertising, and promotions, as well as creating new digital solutions for businesses. 

The Project in Numbers

This project is one of the first where Respeecher worked with such a large volume of data in such a short time.

How was the voice of the broadcasting legend cloned?

First, let's briefly describe how speech synthesis technology works.

In short, you have the voice of the person you want to clone (also called the target voice). You also have the voice of the person that you intend to transform into the target voice (this is called the source voice).

And there is a medium, controlled by AI, which studies the target's voice, understands its features, and then transforms the source's voice into the target's. The services of such a medium are exactly what Respeecher provides.

Respeecher built an audio version of the super resolution algorithm to deliver the highest resolution audio, even in cases when the client doesn't have high-res sources available. Learn more about it by downloading this whitepaper about Respeecher's audio super-resolution algorithm.

This is not the first time we resurrected a sport legend’s voice - we also created a digital Vince Lombardi for the 2021 Super Bowl LV. It’s just that this time we were working under such difficult conditions.

Please see what Vitalii Knyr, the Head of Delivery at Respeecher pointed out as main challenges in this project:

1. Fresh/Experimental model
   When we received the target data we put all our hopes in the new experimental model, which theoretically should have worked better with noisy samples, but we didn't have enough experience with applying it to commercial projects.
2. Language
   Since we didn't have anyone in the team who spoke Spanish, we were not sure if the conversion worked well until we shared the result with DDB. It was a great relief when the client responded "Thank you so much! The sound is perfect!".
3. Zeitnot
   We usually have a couple of days between receiving the final lines to convert and sharing the result with the client, but in this case, it was a couple of hours to do the job on our end. On top of the fact that we didn’t have anyone that could speak Spanish, we only had one shot to do our work well and we couldn't check if it was good or not until the client returned to us the next day.
4. Data quality
   All kidding aside - that was the most complex target data we had ever worked with. The recordings suffered from quality issues of all kinds:
   
   - background noise from the game and audience
   - low sampling rate (16KHz)
   - outdated and low-fi microphones
   information/quality lost when transferring data from analog (VHS, old recorders, etc.) to digital format, and many others.
5. Time Zone difference
   Since the Respeecher team was about 9,000 kilometers away from the client, we had to adjust our working schedule and work overnight to reduce delays in communication and delivery.
   

Thanks to constant improvements to our AI algorithms, we were able to deal with all the challenges. Not only were we able to clone Morales’ voice using antiquated data, but we synthesized his voice for the entire broadcast in just 5 hours!

It was remarkable, not only for ourselves but for our partners from the DDB as well:

"Discovering Respeecher was the highlight of our project. They agreed to work with the task even when our training data was not of the best quality. The result was fantastic, and the level of detail was just outstanding." 

-Joseph Lopez, Marketing Technology Director of DDB Latina

So, we spent 10 days training the AI system's algorithms and 5 hours on the synthesis itself. If you've ever worked in audio production, you can certainly understand how impressive this timeframe is for a 2-hour audio record.

Project's Results and Cultural Impact 

Against the background of discussions about the negative impact of deepfake technologies, it is difficult to overestimate the positive effect of this project.

While working on the project, we not only cooperated with partners from DDB but also with the family of Rivera Morales. For us, this project has become much more than regular work, mainly due to the warm reaction of his family and the people of Puerto Rico.

In comments in the press and on social media, Puerto Ricans also warmly embraced the revival of the commentator's voice from the past. This voice brought people back to a happy time in their youth along with remembering the past successes of the Puerto Rican national team.

This is now one of the hot topics in Puerto Rico and beyond. Respeecher's level of quality and production speed directly states that marketing agencies, filmmakers, and content creators can be more ambitious in their ideas. We have the technology to tackle the toughest challenges.

As for our role in this project, we are always happy to work with history. Resurrecting the voices of historical characters creates a connection between generations separated by time. Past figures can bring a bit of optimism into the present, shake off good memories or make you reflect on some eternal problems.

The potential of our technology for educational and museum projects is genuinely invaluable. We are confident that even more ambitious resurrection projects await us in the future. And that's great because technology should serve the best of our motives.