Khazina run logs — code, losses and evals
The reproducibility record for turning 1,238.9 h of Tajik radio into 860.0 h of verified transcripts, and for the training runs built on it. Code, evals and loss curves only — the audio and the model weights live elsewhere (see below).
What is here
| path | what |
|---|---|
code/ |
every script in the pipeline: download, transcribe, gate, recover, language-check, cut |
eval/ |
the numbers: conversion parity, decode sweep, gate summaries, failure report, language check |
gemma_khazina/ |
the Gemma-4-12B audio stage: MANIFEST, training config, loss.csv |
tts/ |
the 24 kHz TTS set summary and the Kazakh TTS round-trip measurement |
METHOD_AUDIO_TO_TEXT.md |
the playbook: tools, commands, rules, traps, expected yield |
The headline numbers
| input | 2,671 episodes / 1,238.9 h (Khazina radio, TV Tajikistan, presidential channel) |
| speech found | 1,167.3 h |
| verified | 150,553 clips / 860.0 h across 65 books |
| agreement | NormLevDist median 0.095; 86.3% within 0.20 |
| ASR pass 1 | Whisper-large-v3-CA, CT2 fp16, beam 1 — 9.81 WER on the frozen FLEURS-tg judge |
| ASR pass 2 | Parakeet-TDT-0.6b-v3-tg — 11.66 WER, a different architecture |
How the labels were made
No human transcribed this. Every clip was transcribed twice by two models that fail differently — an encoder-decoder and a TDT/CTC — and only clips where both agreed were kept. Disagreements were not deleted; they are recorded with both transcripts and the distance.
The transcripts are pseudo-labels with roughly our ASR error (~9-12% WER). Treat them as such.
Five rules this pipeline is built on
- The second model must be a peer, not a weaker one. A weaker second model measures its own errors: whisper-turbo scored NormLevDist median 0.725 with 0% of clips passing, against Parakeet's 0.107 and 94%.
- Normalise both sides, including numbers. One model writes
77, the otherҳафтоду ҳафт. - Look at the output, not only the metrics. The worst bug produced perfect text with unusable timestamps; every aggregate looked healthy.
- Never guard a stage on its own output file. Twice a stage silently skipped its real work because an output file already existed — once from a race, once from a test run.
- Cap CPU threads per worker. Uncapped, torch gave each of 107 workers ~235 threads on 208 cores; capping at 2 was worth 6.2x throughput and took 8 GPUs from 0-54% to 100%.
Where the rest lives
- 16 kHz audio (ASR):
Tohirju/sl-rowan - 24 kHz audio (TTS):
Tohirju/sl-flint - non-Tajik audio separated out of the corpus:
Tohirju/sl-tundra(Russian),Tohirju/sl-juniper(Kyrgyz)
Licence
Research use only. Saidzoda Lab claims no rights over the underlying recordings. Access is gated and granted manually.
- Downloads last month
- 16