Skip to content
Data collection

Building speech datasets for Indian languages: what it really takes

Hours of audio are easy to count. Useful hours, with the right speakers, conditions and consent, take careful design.

SSkevyr team2 min read

Speech recognition that works for everyone in India has to hear everyone: different languages, accents, ages, devices and rooms. Most public speech data is dominated by a few accents recorded in quiet conditions, which is why voice products often struggle with real users.

#Start with who should be in the data

Before recording a single file, we agree the mix of speakers the model needs: gender, age bands, regions, first language and devices. These become quota cells, and the collection is tracked against each one, so the final dataset isn't quietly skewed towards whoever signed up first.

#Choose the right kind of speech

  • Read speech: speakers read prepared sentences. Easy to transcribe and good for broad coverage of words and sounds.
  • Spontaneous speech: speakers answer questions in their own words. Closer to how people really talk to voice assistants.
  • Conversations: two people talking naturally, including interruptions and code-mixing. Essential for call-centre and dialogue systems.
  • Commands and wake words: short phrases in quiet and noisy rooms, for devices.

#Check quality while recording, not afterwards

Our contributors record in the browser, on their own phones and laptops. Before a session starts, a short room check measures background noise. Each take is checked instantly for length, silence, clipping and noise, so problems are fixed by recording again on the spot instead of being discovered weeks later.

After upload, every file is checked again on the server and converted to the delivery format, and an acoustic fingerprint catches recordings that were submitted twice. Only then does a human reviewer listen and decide.

A rejected recording that gets re-recorded in the same session costs seconds. One discovered at delivery costs the project.

Every contributor agrees, before they start, to how their recordings will be used. Clients receive anonymous speaker codes and aggregate profile information, never names, phone numbers or ID documents. That consent record travels with the dataset, which matters more every year as data-protection rules tighten.

Whether you need a custom collection or want to start from an existing dataset, the same principles apply. Ask us for a sample to hear the difference.

Share this articleLinkedInXWhatsApp
S
Written by Skevyr team

We build training data, expert feedback and evaluations for AI teams. Published 24 September 2026.