[ad_1]
But few people were proficient enough in the language to manually copy the sound. Inspired by voice assistants like Siri, Mahelona began researching natural language processing. “Teaching the computer the Maori language has become absolutely necessary,” says Jones.
But Te Hiku faced a chicken-and-egg problem. build again the speech recognition model needed a lot of replicated voices. To copy the sound, he needed advanced speakers, which he tried to compensate for their small number in the first place. However, there were plenty of beginner and intermediate speakers who could read. again louder than they can recognize words on a recording.
So Jones and Mahelona, along with Te Hiku COO Suzanne Duncan, came up with a clever solution: instead of copying the existing sound, they would ask people to record themselves reading a series of sentences designed to capture all the sounds in the language. According to an algorithm, the resulting dataset will perform the same function. These thousands of couples would learn to recognize from spoken and written sentences. again sound syllables.
The team has declared a competition. Jones, Mahelona, and Duncan contacted every Māori community group they could find. shut up dance groups and waka but He joined the canoe racing teams and announced that whichever submits the most records will win a grand prize of $5,000.
The entire community was mobilized. Competition intensified. Te Mihinga Komene, a Maori community member, is an educator and advocate of using digital technologies to revitalize againalone recorded 4,000 sentences.
Money wasn’t the only source of motivation. People embraced Te Hiku’s vision and relied on him to protect their data. “Te Hiku Media said, ‘Whatever you give us, we are here. kaitiaki [guardians]. We’re interested, but the voice is still yours,” says Te Mihinga. “This is important. These values define who we are as Māori.”
In 10 days, Te Hiku collected 310 hours of transcript pairs from nearly 200,000 recordings made by about 2,500 people; this is an unheard of level of engagement among researchers in the AI community. “No one outside of a Māori organization could have done this,” says Caleb Moses, a Māori data scientist who joined the project after learning about the project from social media.
The amount of data was still small compared to the thousands of hours typically used to train English language models, but it was enough to get started. Using data from the Mozilla Foundation to boot an existing open source model, Te Hiku created its first model. again Speech recognition model with 86% accuracy.
[ad_2]
Source link