Technical Checks
- 44100 Hz sample rate
- mono channel
- 16-bit PCM WAV
- no broken files
A practical deployment checklist for rebuilding the diphone and TTS system safely
Checklist
This page provides the practical rebuild checklist for the Bishnupriya Manipuri diphone and TTS workflow, including rule freezing, folder cleanup, audio preparation, validator testing, and deployment readiness.
A rebuild often fails not because of one big mistake, but because several small mismatches accumulate: changed IPA rules, renamed filenames, leftover old files, inconsistent source audio, or incomplete validation.
Freeze rules ↓ Back up old system ↓ Create clean output folder ↓ Prepare source audio ↓ Generate / segment diphones ↓ Validate ↓ Fix missing or mismatched items ↓ Deploy
Before rebuilding any audio inventory, make sure the logical system is stable. A rebuild should never begin while pronunciation rules are still shifting.
| Rule Layer | Must Be Frozen? | Why |
|---|---|---|
| IPA conversion rules | Yes | Diphone generation depends on stable IPA output |
| Phoneme tokenization | Yes | Tokenization affects diphone sequence structure |
| Diphone generation logic | Yes | Expected transition units must remain consistent |
| Safe filename mapping | Yes | Validator and playback rely on exact file names |
Always back up the old diphone folder and related working files before replacing anything.
Never rebuild by mixing fresh files into an uncertain old folder. Use a new clean directory for the rebuilt output.
old_live_diphone/ ← keep as backup new_rebuild_diphone/ ← use for current rebuild
This avoids confusion caused by leftover files generated under older rule systems.
The rebuild depends on usable word-level source recordings. Before segmentation, check that source audio is clean and technically consistent.
Once source words are prepared, generate expected diphone sequences and segment or extract the corresponding audio units.
Source word ↓ IPA ↓ Phoneme sequence ↓ Diphone sequence ↓ Safe filenames ↓ Segment / save files
Save all new files into the clean rebuild folder, not directly into the live production folder.
Before full deployment, validate against a small but representative word list.
A curated set of 30–50 words is usually enough to expose most system-level mismatches.
The validator should clearly separate passing items from partial or failing ones.
| Status | Meaning | Action |
|---|---|---|
| Pass | All expected files exist and match | No action needed |
| Partial | Some files exist, some are missing | Rebuild only the missing parts |
| Fail | Major mismatch in rule logic or file set | Stop and review rule consistency |
| Rule Mismatch | Expected and generated outputs disagree | Review converter / filename logic before deployment |
Use the validator output as a rebuild target list. Fix only what is wrong, but always check whether the problem is:
After fixing the detected issues, run the validator again on the same test list. If possible, expand testing to a wider list before deployment.
Initial validation ↓ Fix issues ↓ Repeat validation ↓ Confirm stable results ↓ Deploy
Move the rebuilt files into the live folder only after the test set behaves consistently.
| Step | Checklist Item | Done? |
|---|---|---|
| 1 | Freeze IPA conversion rules | ☐ |
| 2 | Freeze phoneme tokenization rules | ☐ |
| 3 | Freeze diphone generation rules | ☐ |
| 4 | Freeze safe filename mapping rules | ☐ |
| 5 | Back up old diphone folder | ☐ |
| 6 | Create clean rebuild folder | ☐ |
| 7 | Verify source audio technical format | ☐ |
| 8 | Trim and normalize recordings | ☐ |
| 9 | Generate expected diphone sequences | ☐ |
| 10 | Segment or prepare diphone files | ☐ |
| 11 | Run validator on curated test list | ☐ |
| 12 | Fix missing or mismatched files | ☐ |
| 13 | Run validator again | ☐ |
| 14 | Confirm word-page playback matches validator | ☐ |
| 15 | Deploy rebuilt folder to production | ☐ |
/audio/diphone_backup_old/ /audio/diphone_rebuild/ /audio/diphone_live/
Keeping backup, rebuild, and live folders separate makes troubleshooting much easier.
Review the source recording and normalization workflow.
Review validator logic, coverage calculation, and failure types.
Review the naming system that must stay stable during rebuild.