Hearthno account · no cap · real Nebius job

Hear 0:14 of voice kept in 3.3 KB.

FLEURS read speech, not a family note: one FLEURS hi_in test sentence (clip 5676772040210847328, the median round-trip drift of 20). Google FLEURS (hi_in test split), CC BY 4.0. The consented, actor-voiced family sample replaces it once recorded.

  1. ▶ Original
  2. ▶ Kept — NanoCodec's reconstruction of the recording
  3. Search “colleagues” — or any word you hear

FLEURS reader

2026-10-15 · 0:14 · hi-IN · real Nebius job aijob-e01ncwsnrzpd08qcv3
0:00tokens…0:14

Kept — NanoCodec's reconstruction of the recording. Decoded on the Nebius Job from the 3.3 KB of tokens drawn above.

Stored3.3 KB for 0:14 of voice
ASR agreement after the round trip97.7 % (an original is deleted only at ≥ 95 %)
Nebius jobaijob-e01ncwsnrzpd08qcv3 · cpu-d3 · 4vcpu-16gb · eu-west1
sha256 of the tokens3b1a340060924b6b…

What was said · Hindi, then machine translation

Kept · quoted as transcribed

No promise or ask in this note passed the gates.

Proposed 0 · kept 0 · dropped by the quote gate 0. The gates prove these words are in the transcript. They do not prove the speech recogniser heard them right — so every coupon says “as transcribed” and plays its source.

Find a moment

815.4 KB per hour of voice|WER 8.7 % → 12.8 % after the round trip (FLEURS hi_in, 20 clips)|RTF ASR 0.16 · codec 0.63|$0.1345 per audio-hour (compute only, list price)|job aijob-e01y33r8hyfcn7np5v · cpu-d3 4vcpu-16gbproof →
Keep your own note

Upload up to 60 seconds in any language the speech model is transcription-ready for (Japanese and Korean excepted); reminders are Hindi-only in v1. It runs a real Nebius Job: measured on this deployment, 6.6 min median from upload to kept (6.7 min at p95, 10 runs), cold start and queue included.

Open an empty space