Hume AI
How Hume AI moves petabytes of audio data without bottlenecks
.png)
Hume AI
Rashish Tandon, VP of Research Infrastructure, Hume AI
egress fees
upload speed
object store
Hume AI trains speech models, which means moving enormous volumes of audio data in and out of storage on a tight timeline. Previously, that data was spread across multiple hyperscale object stores. As Hume's data stores surpassed petabyte scale, the team needed a single, high-throughput home for its data, one that could keep pace as processing pipelines ramped up.
Hume consolidated its data on Backblaze B2 Cloud Storage as one central object store. The team pulls audio from B2 to its compute clusters for processing, then writes new artifacts straight back to B2. For fast IO work, data moves between B2 and an NFS store. Hume's clusters now push close to 10 GB/s, fast enough to keep pace with the team's heaviest processing runs.
Hume now runs its processing, training and evaluation pipelines through a single object store with no one hunting across silos to find data. Upload speeds of close to 10 GB/s mean the team's data keeps pace with its processing, even during heavy ramp periods. That speed matters beyond the infrastructure: Hume's own deliverables to research partners run on tight deadlines, and a storage layer that can't keep up would put those commitments at risk. With Backblaze, it hasn't.
Backblaze B2 sits at the center of Hume's data flow. Raw audio lands in B2, which acts as the permanent home for everything. When Hume needs to process data, it pulls from B2 to its compute clusters, runs the work, then writes new artifacts back to B2. For fast IO jobs, select subsets of data move between B2 and NFS stores. From there, Hume uses the processed data for training, querying, and delivery to its own customers.
.png)
.png){kind=logo}
Hume AI is the data and evaluation layer for emotionally intelligent voice AI. We help teams build and measure voice AI the way people actually experience it - grounded in real human judgment, not just transcripts.
Most of the data Hume keeps in B2 is fresh. As the team ramped up its data efforts, it pushed new audio straight into Backblaze while processing it on compute clusters in parallel, all without slowing the pipeline down. Hume's clusters now sustain close to 10 GB/s of upload speed.
Backblaze has earned our trust by delivering reliable results on the data we have, and so we’re excited to continue to grow our data stores with them.
Rashish Tandon, VP of Research Infrastructure, Hume AI
Hume's data growth does not move on a predictable schedule. Activity can sit quiet for a couple of weeks, then spike fast when a new project needs fresh artifacts. Backblaze's throughput has held up through those swings, which means the team can push hard when it needs to without planning around storage as a bottleneck. That reliability is what convinced Hume to keep expanding what it stores in B2 rather than holding data back in legacy systems.
Our growth doesn't follow a curve, it follows a trigger. When a customer has needs, suddenly we're moving petabytes. Backblaze has kept up every time.
Rashish Tandon, VP of Research Infrastructure, Hume AI
.png){kind=tracking}
