Alert Networks ML Case Study
|Company Name||Alert Networks|
|Case Study Title||Alert Networks Uses Machine Learning Neural Networks to Save Lives|
|Case Study Short Description||Alert Networks provides real-time insights for emergency responders via transcribing audio feeds of emergency radio transmissions. TriumphTech created a transcription solution that streams, up-samples, and transcribes those transmissions at scale.|
|Customer Challenge||Alert Networks’ customers expect real-time insights for emergencies via transcribed emergency radio transmissions. Due to the speed and volume of data, machine learning is required versus human transcription. As well, cost at scale was a concern as pre-existing (AWS Transcribe) services were cost prohibitive. Finally, audio feeds from radios are notoriously noisy and difficult for ML to transcribe. Triumphtech would need to implement a solution that was cost effective and scalable.|
|Proposed Solution and Architecture||Triumphtech proposed three solutions to Alert Networks. First, streaming the data ingestion. This would be accomplished by leveraging Kinesis Data Firehose and Kinesis Video Streams. These services would allow the mass ingestion of data to SageMaker for the second solution.
The second solution was to use ML to upsample the audio to make it easier to transcribe. By using open source and proven Neural Network libraries the radio transmissions could be brought up to a quality that was decent enough for transcription.
The third solution was Triumphtech developing a custom ML model to transcribe the audio to text. This was done with Wav2Ver2. Wav2Vev2 was chosen because it was pre-trained to translate English to test, it could be tuned with custom vocabularies, and AWS uses Wav2Vec2 as the basis for other core AWS services such as Alexa. As such, it integrates well with SageMaker.
|Outcomes of Project & Success Metrics||After multiple deep discoveries with Alert Networks, TriumphTech opted for the above solution based on cost to scale. By using the above three solutions, TriumphTech was able to craft an infrastructure whose costs scaled linearly even with exponential business growth.|
|Date Entered into Production||July 2022|
|Lessons Learned||Transcribing data with machine learning requires a knowledge of the underlying problems that can result from the source of that data.|
|Summary Of Customer Environment||Cloud environment is native cloud. The entire stack is running on Amazon Web Services. Stack is being deployed in the US-East-1 region.|
View more articles View more articles