What actually ran.
Job IDs, presets, timings, bytes and hashes — read from the Jobs' own reports and the bucket, never typed in.
Benchmark of record
| Run | 20261014T231930Z-4vcpu-16gb · Nebius job aijob-e01y33r8hyfcn7np5v · cpu-d3 4vcpu-16gb |
|---|---|
| Stored size | 815.4 KB per hour of voice · 194.7× smaller than 22.05 kHz PCM |
| WER, FLEURS hi_in | 8.7 % original → 12.8 % after the codec round trip (20 clips; strict normaliser 11.2 % → 15.1 %) |
| Real-time factor | ASR 0.165 · codec encode+decode 0.630 · whole pipeline 1.019 |
| Cold start | job created → container running 164 s · install 17 s · both models loaded 153 s · first note processed 349 s after create |
| Cost | $0.1345 per hour of voice: compute only (measured real-time factor × cpu-d3 list price); Job start-up, install and model load excluded |
| Runtime | EncDecRNNTBPEModelWithPrompt on NeMo · att_context_size [56, 13] = 1.12 s chunks · ASR input 16000 Hz · codec 22050 Hz, 1764 samples per frame · 4 torch threads · weights from the in-region bucket |
| Versions | torch 2.13.0+cpu · nemo 3.0.0 · transformers 5.19.0 · numpy 2.4.6 · librosa 0.11.0 · soundfile 0.14.0 · huggingface_hub 2.2.0 · lightning 2.4.0 |
| Codec on CPU | encoding the same input twice gives identical codes: yes · the stored .hct equals a fresh encode: yes · decoding twice, largest sample difference: 0 |
FLEURS is read speech recorded for ASR research, not phone voice notes, and FLEURS is in the ASR's training mix (card), so it is in-domain for this ASR.
The judge sample
| Note | Nebius job | Audio | Tokens | sha256 | Agreement | Pipeline raw copy |
|---|---|---|---|---|---|---|
| one FLEURS hi_in test sentence (clip 5676772040210847328, the median round-trip drift of 20) | aijob-e01ncwsnrzpd08qcv3 | 0:14 | 3.3 KB | 3b1a34006092… | 97.7 % | raw HEAD 404 · originals HEAD 404 2026-10-15T00:17:12.985Z |
| a 60-second note of FLEURS hi_in test sentences, as a 16 kbps Opus voice note | aijob-e01ncwsnrzpd08qcv3 | 1:00 | 13.5 KB | 966fb5b814cf… | 85.3 % | raw HEAD 404 · originals HEAD 404 2026-10-15T00:17:17.253Z |
What “deleted” rests on, checked against the bucket on 2026-10-15: versioning is off, so a delete is final; a DELETE is followed by HEAD 404; and a write with a stale ETag is refused (412) — every lock and merge in Hearth relies on it (evidence/storage-preflight.json).
Memory for the sample: nvidia/nemotron-3-super-120b-a12b on Nebius Token Factory proposed 0 item(s), kept 0. Requests: chatcmpl-807d34a81090affe (0.7 s) · chatcmpl-a8b1e40724d5ca39 (1.2 s).
The memory gates on a live Token Factory call
A test transcript written from the judge sample's script (text, no audio, anchored on Monday 12 Oct) went to nvidia/nemotron-3-super-120b-a12b with a strict json_schema — request chatcmpl-86b6920b56793fdf, 5.8 s. Super proposes; code decides what is shown:
| Super proposed | Quote gate | Amount | Date |
|---|---|---|---|
| «रविवार से पहले पाँच हज़ार भेज देना» · money · ₹5,000 · before sunday | verbatim in the transcript | ₹5,000, parsed from the quote | before Sunday → Sat 17 Oct |
| «उसे फ़ोन कर लेना» · call · on thursday | verbatim in the transcript | not stated | removed: not in the quote → date unclear |
Your last run
Reading your space…
Served by commit 51b3aca.