Click the logo to open the app in UCloud.
When submitting a Transcriber job, we recommend using the gpu-nvidia-b200-1-mig.1g machine type.
In our benchmark, 1 MIG processed a 57 min 48 sec interview in 7 min 20 sec, while an 8-core CPU job took around 1 h 40 min for the same file. Since processing time can vary depending on file type, audio quality, and settings, we use a conservative estimate of 10 minutes of processing time per 1 hour of audio on 1 MIG machine.
As a rough planning estimate, use:
| Machine type | Estimate |
|---|---|
| gpu-nvidia-b200-1-mig.1g | 0.03 GPU-hours per audio hour |
| cpu-amd-zen5-8-vcpu | 15–20 CPU compute-hours per audio hour |
Since GPU-hours must be requested as whole numbers, use the table below when applying for resources:
| Total audio/interview hours | GPU-hours to request |
|---|---|
| 1–33 hours | 1 GPU-hour |
| 34–66 hours | 2 GPU-hours |
| 67–99 hours | 3 GPU-hours |
| 100–132 hours | 4 GPU-hours |
For CPU jobs, use at least 8 cores and around 24 GB memory, for example cpu-amd-zen5-8-vcpu, to reduce the risk of freezing or running out of memory. In our benchmark, RAM usage was around 8.5 GB throughout the process.
For GPU jobs, 1 MIG machine (i.e., gpu-nvidia-b200-1-mig.1g) is recommended. It already provides fast transcription performance and enough GPU memory, thus there is no need to use 2 MIG machines (i.e., gpu-nvidia-b200-2-mig.1g) or a full GPU for Transcriber jobs.
Use the Batch version when possible to avoid wasting resources, as the job stops automatically when transcription is complete. If you use the Default version, click “Show settings” and enable “Transcribe and stop” before starting the transcription.
When the job has completed, the transcription output can be retrieved from the UCloud job page. Select the relevant project workspace, then your completed job and find the output section near the bottom of the page, below “Your job has completed”. The output can also be found in your UCloud files.
