# Mahjong voices v2 Each gender folder contains 51 semantic MP3 filenames. These replace the active voice selection without deleting the older pack. ## Original selected recordings 78 recordings (39 per gender) are preserved byte for byte. Source commit trees supply Git blob SHA-1 verification; `manifest.json` also records final SHA-256, file size, decoded duration and loudness statistics for every clip. - Main set: https://github.com/xiyoufang/mahjong/tree/98080a83759ca17c3129133f5eef373cbe63a814/Resources/raw/Mahjong - Chow: https://github.com/niuma-wj/client/tree/10de6164f7ecbce68dcdcf6f306903bb27d2e108/Assets/NiuMa/Resources/Sounds/Game/Mahjong/Mahjong - Correct mapping: `mjt1_N` is 筒 (`dN`), `mjt2_N` is 万 (`wN`), and `mjt3_N` is 条 (`bN`). - `mjt4_1..7` maps to east, south, west, north, red, green, white. - `peng`, `gang`, `hu`, `zimo`, and `chi` map to `op_pong`, `op_kong`, `op_win`, `op_selfdraw`, and `op_chow`. Public source availability does not establish a recording license. Audio-specific rights were not verified; do not represent these recordings as newly created or freely licensed. The pinned xiyoufang repository tree contains no LICENSE/COPYING file. NiuMa's root MIT license is retained in `NIUMA-MIT-LICENSE.txt`; it grants rights for software and associated documentation but does not independently prove provenance or ownership of third-party voice recordings. ## Newly generated supplements 24 clips (12 per gender) were generated from the explicit Chinese text in the manifest: 春、夏、秋、冬、梅、兰、竹、菊、暗杠、流局、补花、飞龙. - Engine: Microsoft Edge online neural speech through the `edge-tts` Python client. - Female preset: `zh-CN-XiaoxiaoNeural`. - Male preset: `zh-CN-YunxiNeural`. - Rate: +10%; default pitch; short isolated calls. - Processing: silence trim, 80 ms trailing pad, -18 LUFS / -2 dBTP target, mono 24 kHz 96 kbps MP3. - These are distinct generic preset voices, not clones of the original speakers. Matching source timbre or speaker identity is neither attempted nor claimed. A perceptible timbre difference is expected. - `edge-tts` client license: LGPL-3.0 (https://github.com/rany2/edge-tts). This licenses the client, not Microsoft's service or generated audio. Microsoft service/output terms apply; no CC0 or unrestricted commercial-rights claim is made. Terms: https://www.microsoft.com/servicesagreement ## Reproduction and verification Run `python scripts/prepare-voices-v2.py` with `edge-tts` and `imageio-ffmpeg` installed from PyPI. Network access is needed for pinned public originals and preset generation. Existing verified originals/generated outputs are reused; final MP3 decoding and manifests are regenerated. The original selected recordings are not normalized or converted. Verification establishes source identity, coverage, decoding and signal statistics. It does not constitute a human listening test or a legal rights determination. Missing supplements do not reuse the older Kokoro female pack. ## Measured validation results - Total: 102 successfully decoded MP3s, 2,040,847 bytes; 51 per gender. - Female originals: duration 0.43–1.60 seconds, median 0.69; median mean volume -11.9 dBFS; sample peaks -9.4 to 0 dBFS. - Female generated: duration 0.26–0.51 seconds, median 0.325; median mean volume -15.9 dBFS; sample peaks -6.4 to -2.4 dBFS. - Male originals: duration 0.37–1.94 seconds, median 0.52; median mean volume -20.3 dBFS; sample peaks -10.5 to -6.4 dBFS. - Male generated: duration 0.23–0.56 seconds, median 0.415; median mean volume -18.05 dBFS; sample peaks -6.2 to -2.4 dBFS. The supplements have intentionally brief timing; eight of twelve are one-syllable flower names, so duration medians are not an equivalent-text comparison. Source files differ in loudness. Optional app-level attenuation of original female clips to 0.63 and generated male clips to 0.77 approximates within-gender median loudness without re-encoding originals or amplifying peaks. This is approximate signal matching, not perceptual speaker matching.