All learning paths Learn See Try Make Record speech worth keeping What this lesson is about Fix the room, the distance and the fan before the microphone. Most bad audio is made before recording starts.
An hour of Awadhi to record in a shared house at four in the afternoon, with a fan running. Where do you start?
A Record it now, and take the fan noise out of the file afterwards. B Most bad audio is made before the microphone is switched on. C Hold the phone as close to her mouth as you can, and begin.
Check my recall Why this matters IndicVoices records on the speaker's own phone. Each clip is checked at once: audible through the noise, on the prompt, long enough. Fixing the room costs nothing; fixing bad audio afterwards is not possible.
Set up a room and a phone so the speech you record is worth keeping.
IndicVoices records on the speaker's own phone. Each clip is checked at once: audible through the noise, on the prompt, long enough. Fixing the room costs nothing; fixing bad audio afterwards is not possible.
Most bad audio is made before the microphone is switched on. One speaker at a time, ten quiet seconds first, and a listen-back after clip one. Never record over a clip. A bad clip marked bad beats a missing one. Task Recording an hour of Awadhi speech in a shared house at four in the afternoon.
Weak approach Recording straight away, with the fan on and two people talking over each other.
Stronger approach Fan off, phone a hand's width away, one speaker, ten silent seconds, then a listen-back before clip two.
Why the stronger approach works Nothing here needs money. Every one of these is decided before the microphone is on, which is the first rule.
Try a changed situation You are no longer recording voices. You are filming one person's hands doing a task for a robot team. What do you do before the second take?
Make something yourself Record five clips and write a note on the room, the distance and the noise in each. Keep one poor clip, mark it poor, and say what went wrong.
Lesson 7 of 12 on Bharat-language data craft next Check a batch before it goes Part 3 ends in a work sample you can send
Read the primary or official source: Javed et al., IndicVoices (arXiv)