62 lines
5.3 KiB
Markdown
62 lines
5.3 KiB
Markdown
# Embedded announcement audio evidence
|
|
|
|
## Current accepted source
|
|
|
|
Android version 1.5.1 reuses the queue announcement WAV fragments that already existed in this repository's browser implementation. They were first added in commit `6d911271e6bf5b0120edc04c91506077239e96e5`, moved byte-for-byte into `web/src/assets/queue-voice/` in commit `4405bf96f7ad8029fbf62741d92ddaab7b99c096`, and later removed from the browser bundle in commit `5ad8bfcd0377e6dfb8c703d0834d248633a5b939` when that implementation changed to `speechSynthesis`.
|
|
|
|
The user confirmed that the project already had these WAV assets; that evidence redirected the fix to reuse the project's own prior assets instead of generating another voice. Repository history does not record the original generator, speaker, recording author, or a separate license/NOTICE. Therefore this record treats them as project-owned legacy assets in scope for this repair; it does not claim that they are open source or independently cleared for third-party redistribution.
|
|
|
|
## Android mapping
|
|
|
|
| Android resource | Historical source / operation | Spoken content |
|
|
| --- | --- | --- |
|
|
| `voice_prompt.wav` | `please.wav` | 请 |
|
|
| `voice_digit_0.wav` ... `voice_digit_9.wav` | corresponding `digit-0.wav` ... `digit-9.wav` | 零 ... 九 |
|
|
| `voice_range.wav` | PCM concatenation of `number.wav` then `through.wav` | 号至 |
|
|
| `voice_suffix.wav` | PCM concatenation of `number.wav` then `entrance.wav` | 号，前往入口 |
|
|
|
|
The two concatenated resources contain one RIFF header and the source PCM samples in the listed order. Android creates its repeat and batch gaps, so the historical `gap.wav` is not packaged.
|
|
|
|
## Format and integrity
|
|
|
|
All 13 Android resources parse as RIFF PCM format tag 1, 16-bit, mono, 16 kHz, with aligned sample data. `EmbeddedVoiceAssetsTest` locks the exact resource set and validates every committed file through the production WAV parser.
|
|
|
|
| Resource | SHA-256 |
|
|
| --- | --- |
|
|
| `voice_digit_0.wav` | `DA7E001528495F75473F99A6BBC275B9C4B7E5FC0888FA9495F86E16231DDC61` |
|
|
| `voice_digit_1.wav` | `2248F327AC9386ECEADB98A4A24F857C218596406EE519F268D67FEEF785971D` |
|
|
| `voice_digit_2.wav` | `DFCD91AA933F78E5CC4BB18B84A9F00D5365C7992B15EB7948960BFBE4505EE6` |
|
|
| `voice_digit_3.wav` | `98FD18056AF18D3B951E6777D5CA2EE7971005F9FBFD5054E8DADDDA4913778F` |
|
|
| `voice_digit_4.wav` | `073B17DCC81F416799644A34A75723CE1B75A8D90D06679D4F0E6F96DB9CBBFE` |
|
|
| `voice_digit_5.wav` | `80DFA4AD87E752CE23F49B123D91E9E06F691267CBD3ECA94FF12213C73739D7` |
|
|
| `voice_digit_6.wav` | `131A747A9D0C8825F8C69907E19E9FB83935D2EB6FEB1137F6EF0311F96F197E` |
|
|
| `voice_digit_7.wav` | `7DAF943934080733D7F6704CE58FCADC736965A578DC0A091CBFDCCB8E4B9BE4` |
|
|
| `voice_digit_8.wav` | `1406CF912AB2756A6A6600749767FFA0181832DADB28FA0CD81FBB0F1BE47A48` |
|
|
| `voice_digit_9.wav` | `BE72EC45438D4D2835E11B91380715CE36650CC465D5EE5465B540B1B8686F8F` |
|
|
| `voice_prompt.wav` | `C066CAF5107994D4B7F7E96E6818AF54E8BB785914FD5881FA223C18C3CF9BDE` |
|
|
| `voice_range.wav` | `9018BF8EF743F6A6FC2323BDE7313CCFE12D5E14A19207E87C2EE3B3A19D9C7B` |
|
|
| `voice_suffix.wav` | `010CF52AE95B0F541A105DB43741832A75FB670A3D02A72A032D94D9AC95182E` |
|
|
|
|
## Semantic diagnosis and acceptance
|
|
|
|
The 1.5.0 eSpeak NG `cmn` assets were rejected after the target user reported that they did not sound Chinese. Reproduction showed correct byte order and repetition, but two Chinese recognizers returned unrelated text at approximately 0.001-0.002 confidence. Direct unfragmented synthesis from the same engine also failed, confirming the source voice mapping rather than Android composition as the cause.
|
|
|
|
The restored legacy assets were evaluated in the actual Android token order:
|
|
|
|
- `00017`, one pass: `请 + 0 + 0 + 0 + 1 + 7 + 号 + 前往入口`.
|
|
- `00017-00020`, one pass: `请 + 0 + 0 + 0 + 1 + 7 + 号至 + 0 + 0 + 0 + 2 + 0 + 号 + 前往入口`.
|
|
- Android repeats the pass three times with two 350 ms zero-PCM gaps.
|
|
|
|
Vosk small-cn 0.22 recovered the complete target phoneme sequence and sentence structure from final Android-composed samples, with average word confidence approximately 0.84-0.89. Its text used homophones/segmentation such as `一期` for `一七` and `好` for `号`, while all three repeats remained stable. Windows zh-CN SAPI also recovered the sentence anchors but was less accurate on isolated digit fragments. A digit calibration sequence recovered 零 through 九 in order. These are calibrated automated red-flag checks, not a substitute for human acceptance.
|
|
|
|
The final 1.5.1 single-number preview produced from the committed Android resources is 11.516125 seconds and has SHA-256 `9FFDA2837EC40DAF4AC626FF093DDC64223A6F9E352FA3F8E198A48C97C285D3`.
|
|
|
|
## Verification and remaining risk
|
|
|
|
- `gradlew.bat clean testDebugUnitTest lintDebug assembleDebug`: PASS; 53 JVM tests, no failures/errors/skips; lint has zero errors and 23 warnings.
|
|
- APK version: versionCode 8, versionName 1.5.1.
|
|
- APK contains exactly the 13 required `res/raw/voice_*.wav` resources.
|
|
- APK SHA-256: `5CB13105178EA5630EA124D0BE98A189E1B7F0557B2672C8CFAD3B6002FCBBEB`.
|
|
|
|
Remaining acceptance: listen to representative single and range announcements on the target television and confirm every digit. The historical clips have uneven source loudness; they were kept byte-exact rather than normalized because processing would invalidate their prior use and requires renewed listening/ASR review.
|