Practical posts for teams building speech and AI models: how much data you need, where to get it, and what to check before you buy. Written by the people who collect it.
How much audio you need to fine-tune Whisper or another ASR model: what the Whisper paper shows, rough starting points by goal, and how to size a speech dataset without guessing.
Read the postNew posts sit next to our guides, which answer one planning decision each. Pick a topic to narrow the list.
DatasetsWhen to license an off-the-shelf speech dataset and when to commission a custom collection: a side-by-side comparison, five questions that decide it, and what to check in any licence.
Scripted or spontaneous speech, who should be speaking, recording conditions and metadata, plus a brief template you can copy.
Read the guideVerbatim vs clean text, timestamps, speaker labels, and how to decide which one your project needs.
Read the guideThe questions worth asking a vendor, what good answers sound like, and how we answer them ourselves.
Read the guideGuideline clarity, inter-annotator agreement, how edge cases get resolved, and the review layers that catch mistakes.
Read the guideWhat each kind of data is good at, where synthetic data goes wrong, and how to test a model that uses both.
Read the guideThe decisions behind a first-person video dataset, from device and mount to privacy, labels and checks, with a brief template to copy.
Read the guideFollow the RSS feed, or ask to be on the list and we'll email you when a new post goes up. Nothing else.
Ask us. Krishsinh will write back within one business day.