Add align/recheck/compare, plugins, and refresh Agent docs.
selftest / selftest (push) Canceled after 0s

Wire deliver_report/listen_pack/retry hooks; gen --auto-retry and intro light check; make AGENTS_MIN the up-to-date agent entry with full command map.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
swl
2026-10-10 13:03:44 +08:00
co-authored by Cursor
parent b1afe68f3c
commit 6d79ab93a2
20 changed files with 780 additions and 91 deletions
+5 -5
View File
@@ -7,10 +7,10 @@ alwaysApply: true
# songforge # songforge
1. **先读** `AGENTS_MIN.md`;不要默认加载整份 `AGENTS.md` / `ironlaw.py` / `qc_worker.py`。 1. **进仓先读** `AGENTS_MIN.md`(唯一默认入口)。不要默认加载整份 `AGENTS.md` / `ironlaw.py` / `qc_worker.py`。
2. 做歌流水线:`doctor_all` → `target` → 写词 → `check` → `gen` → `qc --baseline`。结束后按 `AGENTS_MIN.md`「生歌报告」模板向用户汇报。 2. 流水线:`doctor_all` → `target` →(可选 `align`)→ `check` → `gen`(可 `--auto-retry`)→ `recheck` → `qc --baseline` →(可选 `compare`)。结束后按 MIN「生歌报告」汇报;优先看 run 目录 `report.md` / `listen/`。
3. 音频探测 / 转码 / 截取:只用 `songforge media info|convert|clip`(经 qc_python)。 3. 音频只用 `media` / `recheck` / `qc`(经 qc_python)。扩展写 `plugins/`,见 `plugins/README.md`。
4. 缺 `local.toml`、Comfy 不通、风格/歌名不清、想改模型或 QC 阈值 → **先问用户**。 4. 缺 `local.toml`、Comfy 不通、风格/歌名不清、想改模型或 QC 阈值 → **先问用户**。
5. **改代码铁律**:发现需改仓库代码时,先说明缺口与范围并**征求同意**;建议用户用 git 分支/commit/stash 备份,或新写插件/旁路少动核心;再问是否 commit、是否 push、推哪个 remote。未获许可不得改核心、不得提交/推送。用户同意收尾时再 commit+push。 5. **改代码铁律**:先说明缺口与范围并征求同意;建议 git 备份或新插件;再问是否 commit/push、推哪个 remote。未许可不得改核心、不得提交/推送。
6. 禁止开曲衬词;默认 lean tags;**禁止**在 tags 点名 mm/ooh/嗯 等音节。扩展写 `plugins/`,少改 `core/`/`rules/`。 6. 禁止开曲衬词;默认 lean tags;禁止在 tags 点名 mm/ooh/嗯。
7. 勿提交 `.secrets/`、`config/local.toml`、`.venv/`;勿 `--force` push。 7. 勿提交 `.secrets/`、`config/local.toml`、`.venv/`;勿 `--force` push。
+8 -3
View File
@@ -125,14 +125,19 @@ doctor → target → 写歌词 → check → gen → qc --baseline
```powershell ```powershell
.\.venv\Scripts\python.exe songforge.py doctor .\.venv\Scripts\python.exe songforge.py doctor
.\.venv\Scripts\python.exe songforge.py target --style pop --bpm 104 --dur 180 .\.venv\Scripts\python.exe songforge.py target --style pop --bpm 104 --dur 180
# 按 target 区间写歌词到 input/lyrics/xxx.txt(勿写 [intro] / tags) # 按 target 写词;密度不够用 align(勿写 [intro] / tags)
.\.venv\Scripts\python.exe songforge.py align --style pop --lyrics-file input\lyrics\xxx.txt --bpm 104 --dur 180
.\.venv\Scripts\python.exe songforge.py check --style pop --title 歌名 --lyrics-file input\lyrics\xxx.txt --bpm 104 --dur 180 .\.venv\Scripts\python.exe songforge.py check --style pop --title 歌名 --lyrics-file input\lyrics\xxx.txt --bpm 104 --dur 180
.\.venv\Scripts\python.exe songforge.py gen --style pop --title 歌名 --lyrics-file input\lyrics\xxx.txt --bpm 104 --dur 180 .\.venv\Scripts\python.exe songforge.py gen --style pop --title 歌名 --lyrics-file input\lyrics\xxx.txt --bpm 104 --dur 180 --auto-retry 2
.\.venv\Scripts\python.exe songforge.py recheck --run <run_id>
.\.venv\Scripts\python.exe songforge.py qc --run <run_id> --baseline .\.venv\Scripts\python.exe songforge.py qc --run <run_id> --baseline
.\.venv\Scripts\python.exe songforge.py compare --runs <a>,<b>
.\.venv\Scripts\python.exe songforge.py catalog --limit 20 .\.venv\Scripts\python.exe songforge.py catalog --limit 20
``` ```
- `check` 不过 → **只改歌词** 插件产物:`report.md`、`listen/`、`intro_check.json`、`retry_hint.json`(见 `plugins/README.md`)。
- `check` 不过 → **只改歌词**(或 `align`)
- `doctor` 服务不可达 → 请用户启动 ComfyUI;核对 `local.toml` - `doctor` 服务不可达 → 请用户启动 ComfyUI;核对 `local.toml`
- `qc` 必须带 `--baseline`;`--run` 可用 run_id 或 `data/runs/...` 路径 - `qc` 必须带 `--baseline`;`--run` 可用 run_id 或 `data/runs/...` 路径
- 索引:`catalog` / `catalog --rebuild` - 索引:`catalog` / `catalog --rebuild`
+65 -52
View File
@@ -1,42 +1,60 @@
# songforge — Agent 极简卡(默认只读这份) # songforge — Agent 极简卡(默认只读这份)
> 省 token:**先读本文**。失败 / 改铁律 / 深挖时再打开 [`AGENTS.md`](AGENTS.md)。 > 省 token:**进仓先读本文**。失败 / 改铁律 / 深挖再打开 [`AGENTS.md`](AGENTS.md)。
> **禁止**把 `ironlaw.py` / `qc_worker.py` 整文件塞进上下文;用 CLI 输出代替。 > **禁止**整文件粘贴 `ironlaw.py` / `qc_worker.py`;用 CLI 输出代替。
## 进仓 30 秒
```powershell
cd <songforge仓根>
.\scripts\doctor_all.ps1
.\.venv\Scripts\python.exe songforge.py --help
```
本机路径:`config/local.toml`(缺则从 `.example` 复制)。算法:`config/default.toml`。
## 交付 ## 交付
硬条件:`gen` 的 run_id + flac + `qc --baseline` → **PASS**。 硬条件:`gen` 的 run_id + flac + `qc --baseline` → **PASS**。
`style_fam` / `filler_ratio` 只是弱参考;**不要**对外说「风格已验证」。 `style_fam` / 全曲 `filler_ratio` 弱参考;**开曲衬词**(`intro_filler` / `recheck`)是硬门。
**不要**说「风格已验证」。
生歌结束后,用下面模板向用户汇报(缺项写 `—`,FAIL 时停在「结论」并说明下一步): 生歌结束后用此模板汇报(缺项写 `—`):
```text ```text
## 生歌报告 ## 生歌报告
- 结论:PASS | FAIL - 结论:PASS | FAIL
- 歌名 / 风格 / BPM / 设定时长:… - 歌名 / 风格 / BPM / 设定时长:…
- run_id:… - run_id:…
- 音频路径:…(flac;可用 media info 补实测时长) - 音频路径:…(flac;可用 media info)
- 歌词:路径 + check 密度结果(音节/区间) - 歌词:路径 + check / align 密度
- QC:verdict;时长;实测 BPM(±18%);voice_nn;开曲衬词(intro_filler,硬门);filler_ratio(全曲弱参考) - QC:verdict;BPM;voice_nn;开曲衬词;filler_ratio(弱)
- 弱参考(勿当闸门):style_fam / style_conf;全曲 filler_ratio - 产物:report.md;listen/(PASS 后);intro_check.json
- 模型/采样:ckpt+LoRA(0.5/0.5);shift/steps/cfg/temperature(应与铁律一致) - 建议:听 listen/;FAIL 换 seed(--auto-retry)或改词
- 种子 / 备注:…(重跑或异常时)
- 建议:听哪一段;是否重 gen / 只改词;是否 commit(若改过仓)
``` ```
## 流水线 插件会在 run 目录自动写 `report.md`(`deliver_report`)。
## 流水线(照这个跑)
```text ```text
doctor_all → target → 写歌词 → check → gen → qc --baseline doctor_all → target → [align] → check → gen [--auto-retry N] → recheck → qc --baseline → compare?
``` ```
仓根:
```powershell ```powershell
.\scripts\doctor_all.ps1 .\.venv\Scripts\python.exe songforge.py target --style pop --bpm 104 --dur 180
.\.venv\Scripts\python.exe songforge.py <命令> # 密度不够时一键对齐(复用已有行,不造新词):
.\.venv\Scripts\python.exe songforge.py align --style pop --lyrics-file input\lyrics\xxx.txt --bpm 104 --dur 180
.\.venv\Scripts\python.exe songforge.py check --style pop --title 歌名 --lyrics-file input\lyrics\xxx.txt --bpm 104 --dur 180
.\.venv\Scripts\python.exe songforge.py gen --style pop --title 歌名 --lyrics-file input\lyrics\xxx.txt --bpm 104 --dur 180 --auto-retry 2
.\.venv\Scripts\python.exe songforge.py recheck --run <run_id>
.\.venv\Scripts\python.exe songforge.py qc --run <run_id> --baseline
.\.venv\Scripts\python.exe songforge.py compare --runs <run_a>,<run_b>
``` ```
仓根入口:`.\.venv\Scripts\python.exe songforge.py <命令>`
嵌套 ComfyUI 时:`cd <ComfyUI>\user\default\workflows` 后 `python songforge\songforge.py …`
## 硬锁定 ## 硬锁定
| 项 | 值 | | 项 | 值 |
@@ -44,57 +62,52 @@ doctor_all → target → 写歌词 → check → gen → qc --baseline
| ckpt+LoRA | `ace_step_1.5_turbo_swl_fused` + `swl_voice_best_ep257`(0.5/0.5) | | ckpt+LoRA | `ace_step_1.5_turbo_swl_fused` + `swl_voice_best_ep257`(0.5/0.5) |
| 采样 | shift=5.5 / steps=28 / cfg=2.0 / temperature=0.68 | | 采样 | shift=5.5 / steps=28 / cfg=2.0 / temperature=0.68 |
| 时长 | 150–200s;8 小节器乐前奏(CLI 注入) | | 时长 | 150–200s;8 小节器乐前奏(CLI 注入) |
| 衬音 | **禁止开曲衬词**(嗯/呀/ooh…);默认 lean tags;QC 前奏窗衬词硬 FAIL | | 衬音 | 禁止开曲衬词;默认 lean tags;禁点名 mm/ooh/嗯 |
| tags | `prefer_lean_tags=true`;勿 `--full-tags`(易诱发开曲衬词);禁点名音节 | | tags | `prefer_lean_tags=true`;勿随意 `--full-tags` |
本机路径:`config/local.toml`。算法:`config/default.toml`。 ## 命令速查(Agent 常用)
## 微内核目录 | 命令 | 用途 |
|------|------|
| `doctor` / `doctor_all.ps1` | 体检 |
| `target` | 查字数窗口 |
| `align` | 歌词音节对齐放行窗 |
| `check` | 本地校验 |
| `gen` | 生成;`--auto-retry N` 开曲 FAIL 换 seed;`--skip-recheck` 跳过轻检 |
| `recheck` | 开曲轻检(前奏窗 ASR) |
| `qc --baseline` | 全量验收(必带 baseline) |
| `compare --runs a,b` | 多 run 对比 |
| `media info\|convert\|clip` | 音频窄接口 |
| `catalog` / `prune` | 索引 / 清 tmp |
## 目录(微内核)
| 层 | 路径 | | 层 | 路径 |
|----|------| |----|------|
| 内核 | `core/`(引擎/路径/workflow)、`rules/`、`cli/` | | 内核 | `core/` `rules/` `cli/` `lyrics/` |
| 插件 | **`plugins/`**(钩子总线 + `builtin/` + 第三方) | | **插件** | **`plugins/`** ← 扩展写这里 |
| 数据 | `input/lyrics/`、`data/runs/`、`config/` | | 数据 | `input/lyrics/` `data/runs/` `config/` |
| 文档 | **本文** → `AGENTS.md` → `docs/使用手册.md` → `plugins/README.md` |
扩展优先写 `plugins/<name>.py` 或 `plugins/<name>/plugin.py`,少改内核。说明见 [`plugins/README.md`](plugins/README.md)。 内置/包式插件:`intro_guard`、`deliver_report`、`listen_pack`、`retry_on_intro_fail`。
钩子:`after_assemble` / `after_gen` / `after_qc`。详见 [`plugins/README.md`](plugins/README.md)。
## 音频栈(librosa 等) ## 音频栈
- **已用**(经 `qc_python`,勿塞进轻量 `.venv`):librosa、soundfile、torch、torchaudio、demucs、whisper 经 `qc_python`(勿塞进轻量 `.venv`):**librosa、soundfile、torch、torchaudio、demucs、whisper**。
- CLI:`media` / `qc` 走该栈;不要在 Agent 里手写探测脚本 `media` / `qc` / `recheck` 都走这栈。
## 音频窄接口(勿自写脚本)
```powershell
.\.venv\Scripts\python.exe songforge.py media info --file path.flac
.\.venv\Scripts\python.exe songforge.py media convert --file in.flac --to wav
.\.venv\Scripts\python.exe songforge.py media clip --file in.flac --start 12 --dur 8 --out out.wav
```
(走 `qc_python`,与 qc 同栈。)
## 遇事先问 ## 遇事先问
| 情况 | 动作 | | 情况 | 动作 |
|------|------| |------|------|
| 无 local.toml / Comfy 不通 / 缺模型 | **问用户**路径与是否已启动 | | 无 local.toml / Comfy 不通 / 缺模型 | **问**路径与是否已启动 |
| 风格/歌名/歌词来源不清 | **问** | | 风格/歌名/歌词来源不清 | **问** |
| 想改采样/换模型/放宽 QC | **拒绝**,除非用户书面坚持 | | 想改采样/换模型/放宽 QC | **拒绝**,除非用户书面坚持 |
| secrets / push 哪个 remote | **问**;勿提交 `.secrets`、`local.toml` | | 需改仓库代码 | 先说明范围征求同意;建议 git 备份或新插件;问 commit/push/remote |
| secrets / 推哪个 remote | **问**;勿提交 `.secrets`、`local.toml` |
## 改代码铁律(防意外) 可自行:`check` 密度失败 → `align` 或改词;漏 `--baseline` 补上;`prune --days 7`。
发现现状不满足需求、**需要改仓库代码**时:
1. **先停手**,说明缺口与拟改范围,**征求用户明确同意**后再动文件。
2. **建议用户选一种安全路径**(可并列问):
- git 备份:新分支 / commit / stash / tag,再改;
- 或**新写插件/旁路脚本**,少动核心铁律与默认配置。
3. **问是否 commit、是否 push、推哪个 remote**;未获许可不得提交或推送。
4. 用户同意收尾时再 `commit` +(若许可)`push`;勿擅自 `--force`、勿提交 secrets / `local.toml`。
可自行处理(不必问改代码):`check` 密度失败只改歌词;漏 `--baseline` 补上;`prune --days 7` 清 tmp。
## 样例 ## 样例
+9 -7
View File
@@ -11,7 +11,7 @@
| [`docs/使用手册.md`](docs/使用手册.md) | 人类 | 命令与配置详解 | | [`docs/使用手册.md`](docs/使用手册.md) | 人类 | 命令与配置详解 |
| [`config/local.toml.example`](config/local.toml.example) | 所有人 | 第一次连本机 ComfyUI | | [`config/local.toml.example`](config/local.toml.example) | 所有人 | 第一次连本机 ComfyUI |
> 给 Agent:默认只读 `AGENTS_MIN.md`。缺 `local.toml`、Comfy 没起、风格/歌名不清、想改模型或阈值时——**先问用户**。需改代码时先征求同意,建议 git 备份或新插件旁路,并问清是否 commit/push。音频探测/转码/截取用 `songforge media`,不要读大源码。 > 给 Agent:**进仓只读 [`AGENTS_MIN.md`](AGENTS_MIN.md)**(含 align/recheck/compare、插件与报告模板)。缺 `local.toml`、Comfy 没起、风格/歌名不清——**先问用户**。扩展写 [`plugins/`](plugins/README.md)。
## 一键体检 ## 一键体检
@@ -71,14 +71,16 @@ cd songforge
```powershell ```powershell
.\.venv\Scripts\python.exe songforge.py target --style pop --bpm 104 --dur 180 .\.venv\Scripts\python.exe songforge.py target --style pop --bpm 104 --dur 180
.\.venv\Scripts\python.exe songforge.py check --style pop --title 灯还亮着 --lyrics-file input\lyrics\pop_灯还亮着.txt .\.venv\Scripts\python.exe songforge.py align --style pop --lyrics-file input\lyrics\pop_灯还亮着.txt --bpm 104 --dur 180
.\.venv\Scripts\python.exe songforge.py gen --style pop --title 灯还亮着 --lyrics-file input\lyrics\pop_灯还亮着.txt .\.venv\Scripts\python.exe songforge.py check --style pop --title 灯还亮着 --lyrics-file input\lyrics\pop_灯还亮着.txt
.\.venv\Scripts\python.exe songforge.py qc --run gen-<时间戳> --baseline .\.venv\Scripts\python.exe songforge.py gen --style pop --title 灯还亮着 --lyrics-file input\lyrics\pop_灯还亮着.txt --auto-retry 2
.\.venv\Scripts\python.exe songforge.py prune --days 7 # 清过期 tmp .\.venv\Scripts\python.exe songforge.py recheck --run gen-<时间戳>
.\.venv\Scripts\python.exe songforge.py prune --days 7 --dry-run .\.venv\Scripts\python.exe songforge.py qc --run gen-<时间戳> --baseline
.\.venv\Scripts\python.exe songforge.py compare --runs gen-a,gen-b
.\.venv\Scripts\python.exe songforge.py prune --days 7
``` ```
`qc` 必须带 `--baseline`。样例歌词见 `input/lyrics/`。Agent 规则见 [`AGENTS.md`](AGENTS.md)。 `qc` 必须带 `--baseline`。PASS 后看 `report.md` 与 `listen/`。Agent 入口:[`AGENTS_MIN.md`](AGENTS_MIN.md)。
## 硬约束(摘要) ## 硬约束(摘要)
+57 -4
View File
@@ -340,6 +340,9 @@ def cmd_gen(args):
dry_run=True) dry_run=True)
run.save_json("manifest.json", man_dry) run.save_json("manifest.json", man_dry)
catmod.append_entry(cfg, catmod.entry_from_gen(run.dir, man_dry)) catmod.append_entry(cfg, catmod.entry_from_gen(run.dir, man_dry))
from ..plugins.hooks import bus
bus.emit("after_gen", run_dir=run.dir, manifest=man_dry,
cfg=cfg, args=args)
log.info("dry-run 结束", seeds=seeds) log.info("dry-run 结束", seeds=seeds)
log.close() log.close()
return 0 return 0
@@ -351,8 +354,21 @@ def cmd_gen(args):
if not ok: if not ok:
raise ServerError("ComfyUI 不可达: %s" % info) raise ServerError("ComfyUI 不可达: %s" % info)
from ..plugins.hooks import bus
from .extras import run_intro_check
from ..plugins.retry_on_intro_fail.plugin import should_retry_intro_blob
cro = pmod.comfy_root(cfg=cfg)
music = pmod.music_dir(cfg, cro)
auto_retry = int(getattr(args, "auto_retry", 0) or 0)
do_recheck = not bool(getattr(args, "skip_recheck", False))
jobs = list(built)
results = [] results = []
for sd, wf, wf_path, sub in built: retries_left = auto_retry
ji = 0
while ji < len(jobs):
sd, wf, wf_path, sub = jobs[ji]
ji += 1
pid = eng.submit_retry(wfmod.render(wf)) pid = eng.submit_retry(wfmod.render(wf))
log.info("已提交", seed=sd, prompt_id=pid) log.info("已提交", seed=sd, prompt_id=pid)
@@ -364,7 +380,40 @@ def cmd_gen(args):
results.append(dict(seed=sd, prompt_id=pid, files=outs, results.append(dict(seed=sd, prompt_id=pid, files=outs,
workflow=os.path.basename(wf_path))) workflow=os.path.basename(wf_path)))
cro = pmod.comfy_root(cfg=cfg) if not do_recheck or not outs:
continue
audio = None
for fn in outs:
p = fn if os.path.isabs(fn) else os.path.join(
music, os.path.basename(fn))
if os.path.isfile(p):
audio = p
break
if not audio:
continue
_code, blob = run_intro_check(
cfg, audio, bpm=st["bpm"], intro_bars=ibars, lang=lang)
run.save_json("intro_check.json", blob or {})
log.info("开曲轻检", ok=(blob or {}).get("ok"))
if should_retry_intro_blob(blob) and retries_left > 0:
retries_left -= 1
sd2 = seed_base + len(jobs)
sub2 = "%s/%s-s%d-%s-%s" % (
prefix, safe_title, len(jobs) + 1, style_tag, date_stamp)
wf2 = wfmod.build_workflow(
template, title=title, tags=tags, lyrics=lyrics, seed=sd2,
bpm=st["bpm"], duration=st["duration"], keyscale=st["key"],
sampler=cfg["sampler"], encode=cfg["encode"],
model=cfg["model"], output_prefix=sub2, output_format=fmt,
language=lang,
timesignature=cfg["song"]["time_signature"])
wf_path2 = run.path("wf_%s_s%d.json" % (safe_title, len(jobs) + 1))
with open(wf_path2, "w", encoding="utf-8") as f:
json.dump(wf2, f, ensure_ascii=False, indent=1)
jobs.append((sd2, wf2, wf_path2, sub2))
log.warn("开曲轻检 FAIL,追加 seed 重试", seed=sd2,
left=retries_left)
manifest = dict( manifest = dict(
run_id=run.run_id, title=title, style=args.style, run_id=run.run_id, title=title, style=args.style,
bpm=st["bpm"], key=st["key"], duration=st["duration"], bpm=st["bpm"], key=st["key"], duration=st["duration"],
@@ -375,16 +424,18 @@ def cmd_gen(args):
intro_bars=ibars, intro_bars=ibars,
sampler=cfg["sampler"], encode=cfg["encode"], model=cfg["model"], sampler=cfg["sampler"], encode=cfg["encode"], model=cfg["model"],
density=stats, results=results, density=stats, results=results,
output_dir=pmod.music_dir(cfg, cro), output_dir=music,
created=datetime.datetime.now().isoformat(timespec="seconds"), created=datetime.datetime.now().isoformat(timespec="seconds"),
) )
mpath = run.save_json("manifest.json", manifest) mpath = run.save_json("manifest.json", manifest)
catmod.append_entry(cfg, catmod.entry_from_gen(run.dir, manifest)) catmod.append_entry(cfg, catmod.entry_from_gen(run.dir, manifest))
bus.emit("after_gen", run_dir=run.dir, manifest=manifest, cfg=cfg, args=args)
log.info("全部完成", run=run.run_id, manifest=os.path.basename(mpath)) log.info("全部完成", run=run.run_id, manifest=os.path.basename(mpath))
print("\n生成完成 run_id=%s" % run.run_id) print("\n生成完成 run_id=%s" % run.run_id)
print(" 时长 %ds BPM %d %s 种子 %s" print(" 时长 %ds BPM %d %s 种子 %s"
% (st["duration"], st["bpm"], st["key"], seeds)) % (st["duration"], st["bpm"], st["key"],
[r["seed"] for r in results]))
if inst: if inst:
print(" [纯器乐] 无可唱歌词,密度闸门不适用") print(" [纯器乐] 无可唱歌词,密度闸门不适用")
else: else:
@@ -395,6 +446,8 @@ def cmd_gen(args):
print(" seed=%s -> %s" % (r["seed"], ", ".join(r["files"] or ["无输出"]))) print(" seed=%s -> %s" % (r["seed"], ", ".join(r["files"] or ["无输出"])))
print(" 产物目录: %s" % run.dir) print(" 产物目录: %s" % run.dir)
print(" 音频输出: %s" % manifest["output_dir"]) print(" 音频输出: %s" % manifest["output_dir"])
if os.path.isfile(run.path("report.md")):
print(" 报告: %s" % run.path("report.md"))
log.close() log.close()
return 0 return 0
+196
View File
@@ -0,0 +1,196 @@
# -*- coding: utf-8 -*-
"""扩展命令:compare / align / recheck。"""
import json
import os
import subprocess
from ..config import loader
from ..core import catalog as catmod
from ..core import paths as pmod
from ..lyrics import align as alignmod
from ..lyrics import meter
from .common import build_cfg, resolve_profile, runs_root, style_of
from .qc import find_qc_python
def _load_run(cfg, run_id):
rundir = catmod.find_run_dir(runs_root(cfg), run_id)
if not rundir:
return None, None, None
man = {}
mp = os.path.join(rundir, "manifest.json")
if os.path.isfile(mp):
with open(mp, encoding="utf-8") as f:
man = json.load(f)
qc = {}
qp = os.path.join(rundir, "qc.json")
if os.path.isfile(qp):
with open(qp, encoding="utf-8") as f:
qc = json.load(f)
return rundir, man, qc
def _qc_rows(qc):
if not qc:
return []
return qc.get("rows") or qc.get("results") or []
def _first_audio(cfg, man, rundir):
music = man.get("output_dir") or pmod.music_dir(cfg, pmod.comfy_root(cfg=cfg))
for r in man.get("results") or []:
for fn in r.get("files") or []:
p = fn if os.path.isabs(fn) else os.path.join(music, os.path.basename(fn))
if os.path.isfile(p):
return p
# 也搜 rundir
p2 = os.path.join(rundir, os.path.basename(fn))
if os.path.isfile(p2):
return p2
return None
def cmd_compare(args):
cfg = loader.load({})
ids = []
for part in (args.runs or []):
ids.extend([x.strip() for x in part.split(",") if x.strip()])
if len(ids) < 2:
print("[错误] 至少两个 run:songforge compare --runs a,b")
return 2
print("%-26s %-8s %-5s %-5s %-6s %-8s %-7s %-8s %s"
% ("run_id", "style", "BPM", "dur", "QC", "onset", "衬词比",
"voice_nn", "title"))
for rid in ids:
rundir, man, qc = _load_run(cfg, rid)
if not rundir:
print("%-26s (找不到)" % rid)
continue
rows = [r for r in _qc_rows(qc) if r.get("group") != "baseline"]
r0 = rows[0] if rows else {}
verd = None
for v in (qc.get("verdicts") or []):
verd = v.get("verdict")
break
if not verd:
verd = man.get("qc_verdict") or "-"
print("%-26s %-8s %-5s %-5s %-6s %-8s %-7s %-8s %s"
% ((man.get("run_id") or rid)[:26],
(man.get("style") or "-")[:8],
man.get("bpm") or "-",
man.get("duration") or "-",
(verd or "-")[:6],
r0.get("vocal_onset") if r0.get("vocal_onset") is not None else "-",
("%.0f%%" % (float(r0.get("intro_filler_ratio") or 0) * 100))
if r0 else "-",
(str(r0.get("voice_nn") or "-"))[:8],
(man.get("title") or "-")[:16]))
return 0
def cmd_align(args):
cfg = build_cfg(args)
st = style_of(cfg, args.style, args)
if args.lyrics_file:
with open(args.lyrics_file, encoding="utf-8") as f:
raw = f.read()
elif args.lyrics:
raw = args.lyrics
else:
print("[错误] 需要 --lyrics-file 或 --lyrics")
return 2
prof = resolve_profile(args)
cfg_ly = meter.pick(cfg["lyrics"], prof)
out, diag = alignmod.align_lyrics(
raw, st["duration"], st["bpm"], cfg_ly, target=args.target)
dest = args.out or (args.lyrics_file and (
os.path.splitext(args.lyrics_file)[0] + "_aligned.txt"))
if not dest:
dest = os.path.join(pmod.tmp_today(cfg), "aligned.txt")
pmod.ensure_dir(os.path.dirname(dest))
with open(dest, "w", encoding="utf-8") as f:
f.write(out)
print("== align ==")
print(" 音节 %d -> %d 目标 %d 窗口 %d~%d ok=%s"
% (diag["before"], diag["after"], diag["want"],
diag["win_lo"], diag["win_hi"], diag.get("ok")))
print(" 写出: %s" % dest)
if not diag.get("ok"):
print(" [WARN] %s" % diag.get("note"))
return 1
return 0
def run_intro_check(cfg, audio_path, bpm=104, intro_bars=8, lang="zh"):
"""子进程开曲轻检。返回 (code, blob)。"""
py = find_qc_python(cfg)
if not py:
return 2, dict(error="no qc_python")
worker = os.path.join(pmod.PKG_ROOT, "tools", "intro_check_worker.py")
job = dict(file=os.path.abspath(audio_path), bpm=bpm,
intro_bars=intro_bars, lang=lang)
env = dict(os.environ)
env["PYTHONIOENCODING"] = "utf-8"
env["SONGFORGE_COMFY_ROOT"] = pmod.comfy_root(cfg=cfg)
try:
r = subprocess.run(
[py, worker, json.dumps(job, ensure_ascii=False)],
capture_output=True, env=env, timeout=600)
except subprocess.TimeoutExpired:
return 2, dict(error="timeout")
out = r.stdout.decode("utf-8", "replace")
if "###JSON###" not in out:
return 2, dict(error="no json", log=out[-800:])
blob = json.loads(out.split("###JSON###", 1)[1].strip())
return (0 if not blob.get("error") else 1), blob
def cmd_recheck(args):
cfg = loader.load({})
rundir, man, _qc = _load_run(cfg, args.run)
if not rundir:
print("[错误] 找不到 run: %s" % args.run)
return 2
audio = args.file or _first_audio(cfg, man, rundir)
if not audio or not os.path.isfile(audio):
print("[错误] 找不到音频。可传 --file")
return 2
bpm = args.bpm or man.get("bpm") or 104
bars = args.intro_bars or man.get("intro_bars") or 8
lang = args.lang or man.get("language") or "zh"
code, blob = run_intro_check(cfg, audio, bpm=bpm, intro_bars=bars, lang=lang)
outp = os.path.join(rundir, "intro_check.json")
with open(outp, "w", encoding="utf-8") as f:
json.dump(blob, f, ensure_ascii=False, indent=1)
print(json.dumps(blob, ensure_ascii=False, indent=1))
print("报告: %s" % outp)
if blob.get("error"):
return 2
return 0 if blob.get("ok") else 1
def add_extras(sub):
c = sub.add_parser("compare", help="对比多次 run 的 QC/开曲指标")
c.add_argument("--runs", action="append", required=True,
help="run_id 列表,逗号分隔,可多次")
c.set_defaults(func=cmd_compare)
a = sub.add_parser("align", help="歌词音节一键对齐到 target 放行窗")
a.add_argument("--style", required=True)
a.add_argument("--lyrics-file", dest="lyrics_file")
a.add_argument("--lyrics")
a.add_argument("--out", help="输出路径(默认 *_aligned.txt)")
a.add_argument("--bpm", type=int)
a.add_argument("--dur", type=int)
a.add_argument("--target", type=int, help="目标音节(默认取建议值)")
a.add_argument("--chill-vocal", dest="chill_vocal", action="store_true")
a.add_argument("--density-profile", dest="density_profile")
a.set_defaults(func=cmd_align)
r = sub.add_parser("recheck", help="开曲轻检(前奏窗 ASR 衬词)")
r.add_argument("--run", required=True)
r.add_argument("--file", help="直接指定音频")
r.add_argument("--bpm", type=int)
r.add_argument("--intro-bars", dest="intro_bars", type=int)
r.add_argument("--lang", default=None)
r.set_defaults(func=cmd_recheck)
+8 -1
View File
@@ -3,7 +3,7 @@
命令: 命令:
doctor / styles / target / check / gen / runs / show / catalog / doctor / styles / target / check / gen / runs / show / catalog /
prune / media / qc / selftest prune / media / qc / compare / align / recheck / selftest
设计约束:所有业务参数来自配置/命令行,本文件不写死任何模型或采样常量。 设计约束:所有业务参数来自配置/命令行,本文件不写死任何模型或采样常量。
实现见 common / assemble / commands / selftest / qc。 实现见 common / assemble / commands / selftest / qc。
@@ -120,6 +120,10 @@ def build_parser():
help="注入完整 RULE_NEG(默认关闭;易诱发开曲衬词)") help="注入完整 RULE_NEG(默认关闭;易诱发开曲衬词)")
g.add_argument("--dry-run", dest="dry_run", action="store_true", g.add_argument("--dry-run", dest="dry_run", action="store_true",
help="只写 workflow 与 API 预览,不提交 ComfyUI") help="只写 workflow 与 API 预览,不提交 ComfyUI")
g.add_argument("--auto-retry", dest="auto_retry", type=int, default=0,
help="开曲轻检 FAIL 时额外换 seed 重提次数(默认 0)")
g.add_argument("--skip-recheck", dest="skip_recheck", action="store_true",
help="生成后不做开曲轻检(默认会做并写 intro_check.json)")
g.set_defaults(func=cmds.cmd_gen) g.set_defaults(func=cmds.cmd_gen)
r = sub.add_parser("runs", help="列出历史运行") r = sub.add_parser("runs", help="列出历史运行")
@@ -153,6 +157,9 @@ def build_parser():
from .qc import add_qc from .qc import add_qc
add_qc(sub) add_qc(sub)
from .extras import add_extras
add_extras(sub)
return p return p
+13
View File
@@ -404,6 +404,19 @@ def _render(rep, args, rundir=None, cfg=None):
except Exception as e: except Exception as e:
print("[WARN] catalog 更新失败: %s" % e) print("[WARN] catalog 更新失败: %s" % e)
if rundir and cfg:
try:
from ..plugins.hooks import bus
bus.emit("after_qc", run_dir=rundir, report=rep, verd=verd,
rows=rows, cfg=cfg, args=args)
except Exception as e:
print("[WARN] after_qc 插件: %s" % e)
if os.path.isfile(os.path.join(rundir, "report.md")):
print("交付报告: %s" % os.path.join(rundir, "report.md"))
listen = os.path.join(rundir, "listen")
if os.path.isdir(listen):
print("听感包: %s" % listen)
if not verd: if not verd:
print("本次只有基线校准,没有待验收新歌 -> 退出码 0") print("本次只有基线校准,没有待验收新歌 -> 退出码 0")
return 0 return 0
+21 -3
View File
@@ -778,6 +778,8 @@ def _t_plugins_and_lean_default():
"""微内核钩子已加载;默认 prefer_lean_tags;assemble 不注入 RULE_NEG 堆。""" """微内核钩子已加载;默认 prefer_lean_tags;assemble 不注入 RULE_NEG 堆。"""
from ..plugins.hooks import bus from ..plugins.hooks import bus
assert bus._handlers.get("after_assemble"), "intro_guard 未挂 after_assemble" assert bus._handlers.get("after_assemble"), "intro_guard 未挂 after_assemble"
assert bus._handlers.get("after_gen"), "deliver_report 未挂 after_gen"
assert bus._handlers.get("after_qc"), "qc 插件未挂 after_qc"
cfg = loader.load({}) cfg = loader.load({})
assert cfg["rules"].get("prefer_lean_tags") is True assert cfg["rules"].get("prefer_lean_tags") is True
@@ -791,7 +793,6 @@ def _t_plugins_and_lean_default():
st = {"bpm": 104, "duration": 180, "key": "F major", st = {"bpm": 104, "duration": 180, "key": "F major",
"style": "Chinese pop", "instruments": "kick snare bass", "style": "Chinese pop", "instruments": "kick snare bass",
"mood": "bright", "extra": ""} "mood": "bright", "extra": ""}
# 密度够过闸的短样
raw = ("[verse]\n" + "\n".join( raw = ("[verse]\n" + "\n".join(
["街口的灯还亮着我往前走一步"] * 20)) ["街口的灯还亮着我往前走一步"] * 20))
tags, _ly = assemble(_A(), cfg, st, raw, "selftest") tags, _ly = assemble(_A(), cfg, st, raw, "selftest")
@@ -799,8 +800,25 @@ def _t_plugins_and_lean_default():
"默认 lean 仍注入了 RULE_NEG opening 堆" "默认 lean 仍注入了 RULE_NEG opening 堆"
assert "no wordless vocals" in tags.lower() assert "no wordless vocals" in tags.lower()
assert "mm mm" not in tags.lower() assert "mm mm" not in tags.lower()
return [("微内核与 lean 默认",
"plugins.after_assemble 已挂;默认 tags 为 lean 且无音节点名")] # align:过少歌词应能拉进窗口
from ..lyrics import align as alignmod
from ..lyrics import meter as mmod
thin = "[verse]\n太短了\n再一句\n"
out, diag = alignmod.align_lyrics(thin, 180, 104, cfg["lyrics"])
assert diag.get("ok"), "align 未能把过少歌词对齐: %s" % diag
assert diag["after"] >= diag["win_lo"]
# deliver_report 钩子可写报告
import tempfile
td = tempfile.mkdtemp(prefix="sf-rep-")
man = dict(run_id="t", title="测", style="pop", bpm=104, duration=180,
results=[dict(seed=1, files=["x.flac"])])
bus.emit("after_gen", run_dir=td, manifest=man, cfg=cfg, args=None)
assert os.path.isfile(os.path.join(td, "report.md")), "未写出 report.md"
return [("微内核/lean/align/报告",
"钩子齐;lean 默认;align 进窗;report.md 可写")]
def cmd_selftest(args): def cmd_selftest(args):
+6 -6
View File
@@ -28,12 +28,12 @@ python songforge\songforge.py <命令> [选项]
### 给 Agent / 自动化 ### 给 Agent / 自动化
1. 先读 [`AGENTS_MIN.md`](../AGENTS_MIN.md);不够再读 [`AGENTS.md`](../AGENTS.md)。 1. 进仓先读 [`AGENTS_MIN.md`](../AGENTS_MIN.md)(含 align/recheck/compare/插件);不够再读 [`AGENTS.md`](../AGENTS.md)。
2. 没有 `local.toml`、Comfy 没起、风格/歌名/歌词来源不清 → **先问用户**,不要猜路径或改铁律。 2. 没有 `local.toml`、Comfy 没起、风格/歌名/歌词来源不清 → **先问用户**。
3. 需改仓库代码时:先征求同意;建议 git 备份或新插件旁路(目录 `plugins/`);问清是否 commit/push、推哪个 remote。 3. 需改代码:先征求同意;优先 [`plugins/`](../plugins/README.md);问清 commit/push/remote。
4. 交付以 `qc --baseline` PASS 为准;开曲衬词(`intro_filler`)是硬门;`style_fam` / 全曲 `filler_ratio` 仍是弱参考。 4. 交付:`qc --baseline` PASS;看 `report.md` / `listen/`;开曲衬词硬门;`style_fam` 弱参考。
5. 探测/转码/截取用 `songforge media`(librosa/soundfile 在 `qc_python`),不要整读 `qc_worker.py`。 5. 音频用 `media` / `recheck` / `qc`(`qc_python`:librosa/demucs/whisper…)。
6. 默认 `prefer_lean_tags`;不要随意 `--full-tags`(否定句膨胀会诱发开曲衬词)。 6. 默认 `prefer_lean_tags`;密度不够用 `align`;开曲 FAIL 用 `gen --auto-retry`。
--- ---
+92
View File
@@ -0,0 +1,92 @@
# -*- coding: utf-8 -*-
"""把歌词正文音节对齐到 meter 放行窗口(一键凑密度)。"""
import re
from . import meter
_TAG = re.compile(r"^\[.+\]\s*$")
def _body_lines(text):
"""保留结构:返回 [(is_tag, line)]。"""
rows = []
for ln in text.split("\n"):
s = ln.rstrip("\r")
if _TAG.match(s.strip()):
rows.append((True, s))
else:
rows.append((False, s))
return rows
def _syllables_of_rows(rows):
body = "\n".join(ln for is_tag, ln in rows if not is_tag and ln.strip())
return meter.syllable(body)
def align_lyrics(lyrics, duration, bpm, cfg_lyrics, target=None):
"""调整正文行数使音节落在放行窗;返回 (新歌词, 诊断 dict)。
过少:循环复用已有可唱行(不造新词)。
过多:从末尾删可唱行(保留至少每段一行若可能)。
结构标记与空行尽量保留。
"""
cfg_lyrics = meter.pick(cfg_lyrics, None)
mid, wlo, whi = meter.target_words(duration, bpm, cfg_lyrics)
want = int(target) if target is not None else mid
want = max(wlo, min(want, whi))
rows = _body_lines(lyrics)
before = _syllables_of_rows(rows)
singable = [ln for is_tag, ln in rows if (not is_tag) and ln.strip()]
if not singable:
return lyrics, dict(before=before, after=before, want=want,
win_lo=wlo, win_hi=whi, changed=False,
note="无可唱行")
# 过少:在最后一个非空可唱行后插入副本
guard = 0
while _syllables_of_rows(rows) < wlo and guard < 500:
guard += 1
# 找最后一个可唱行位置
idx = None
for i in range(len(rows) - 1, -1, -1):
if (not rows[i][0]) and rows[i][1].strip():
idx = i
break
if idx is None:
break
donor = singable[guard % len(singable)]
rows.insert(idx + 1, (False, donor))
# 过多:从末尾删可唱行,但至少留 4 行
guard = 0
while _syllables_of_rows(rows) > whi and guard < 500:
guard += 1
n_sing = sum(1 for is_tag, ln in rows if (not is_tag) and ln.strip())
if n_sing <= 4:
break
for i in range(len(rows) - 1, -1, -1):
if (not rows[i][0]) and rows[i][1].strip():
rows.pop(i)
break
# 微调到 want 附近(仍夹在窗口内)
guard = 0
while _syllables_of_rows(rows) < want and _syllables_of_rows(rows) < whi \
and guard < 200:
guard += 1
idx = next((i for i in range(len(rows) - 1, -1, -1)
if (not rows[i][0]) and rows[i][1].strip()), None)
if idx is None:
break
rows.insert(idx + 1, (False, singable[guard % len(singable)]))
after = _syllables_of_rows(rows)
out = "\n".join(ln for _, ln in rows)
if not out.endswith("\n"):
out += "\n"
ok = wlo <= after <= whi
return out, dict(before=before, after=after, want=want,
win_lo=wlo, win_hi=whi, changed=(out != lyrics),
ok=ok, note=("ok" if ok else "仍未落入窗口"))
+31 -9
View File
@@ -1,17 +1,39 @@
# plugins — 扩展目录(微内核外挂) # plugins — 扩展目录(微内核外挂)
Agent / 开发者:**新能力优先写插件**,少改 `core/` / `rules/` / `cli/`。
| 路径 | 用途 | | 路径 | 用途 |
|------|------| |------|------|
| `plugins/hooks.py` | 钩子总线(内核只 `emit`,不 import 业务插件) | | `hooks.py` | 钩子总线(内核只 `emit`) |
| `plugins/builtin/` | 内置插件(随仓发布) | | `builtin/` | 随仓内置(如 `intro_guard`) |
| `plugins/<name>.py` | 单文件插件 | | `<name>/plugin.py` | 包式插件(推荐) |
| `plugins/<name>/plugin.py` | 包式插件入口 | | `<name>.py` | 单文件插件 |
## 钩子 ## 钩子
- `before_build_tags` | 钩子 | 何时 | 典型用途 |
- `after_assemble`(可返回 `{"tags","lyrics"}`) |------|------|----------|
- `before_qc_judge` / `after_qc_judge`(QC 子进程侧另见 qc_worker) | `before_build_tags` | assemble 前 | 改风格参数 |
| `after_assemble` | tags/歌词就绪 | 守卫、改写 |
| `after_gen` | gen/dry-run 写完 manifest | `deliver_report` |
| `after_qc` | qc 渲染结束 | `deliver_report` / `listen_pack` / `retry_on_intro_fail` |
内核层:`core/`(引擎/路径/workflow)、`rules/`(铁律)、`cli/`(命令)。 ## 当前插件
业务扩展优先写插件,少改内核。
| 插件 | 作用 |
|------|------|
| `builtin/intro_guard` | tags 禁止点名衬词音节 |
| `deliver_report` | 写 `data/runs/.../report.md` |
| `listen_pack` | QC PASS 后截 `listen/intro.wav` + `chorus_approx.wav` |
| `retry_on_intro_fail` | 写 `retry_hint.json`;配合 `gen --auto-retry` |
## 最小插件示例
```python
# plugins/hello/plugin.py
from ..hooks import bus
@bus.on("after_gen")
def _hi(run_dir=None, **kw):
print("[hello]", run_dir)
```
+1
View File
@@ -0,0 +1 @@
# package plugin
+84
View File
@@ -0,0 +1,84 @@
# -*- coding: utf-8 -*-
"""生歌结束后按 AGENTS_MIN 模板写 report.md。"""
import json
import os
from ..hooks import bus
def _write_report(rundir, man, qc=None, intro=None):
if not rundir:
return
rows = []
verd = "-"
if qc:
rows = [r for r in (qc.get("rows") or []) if r.get("group") != "baseline"]
vs = qc.get("verdicts") or []
if vs:
verd = vs[0].get("verdict") or "-"
r0 = rows[0] if rows else {}
audio = "-"
for r in man.get("results") or []:
files = r.get("files") or []
if files:
audio = files[0]
break
intro_line = "-"
if intro:
intro_line = "ok=%s n=%s ratio=%s" % (
intro.get("ok"), intro.get("intro_filler_n"),
intro.get("intro_filler_ratio"))
elif r0:
intro_line = "n=%s ratio=%s" % (
r0.get("intro_filler_n"), r0.get("intro_filler_ratio"))
body = "\n".join([
"## 生歌报告",
"- 结论:%s" % (verd if verd != "-" else (man.get("qc_verdict") or "—")),
"- 歌名 / 风格 / BPM / 设定时长:%s / %s / %s / %s" % (
man.get("title") or "—", man.get("style") or "—",
man.get("bpm") or "—", man.get("duration") or "—"),
"- run_id:%s" % (man.get("run_id") or os.path.basename(rundir)),
"- 音频路径:%s" % audio,
"- 歌词:见本目录 lyrics.txt;密度见 manifest.density",
"- QC:verdict=%s;voice_nn=%s sim=%s;开曲衬词 %s;filler_ratio=%s" % (
verd, r0.get("voice_nn") or "—", r0.get("voice_nn_sim") or "—",
intro_line, r0.get("filler_ratio") if r0 else "—"),
"- 弱参考:style_fam=%s style_conf=%s" % (
r0.get("style_fam") or "—", r0.get("style_conf") or "—"),
"- 模型/采样:见 manifest.model / sampler / encode",
"- 种子:%s" % (
", ".join(str(x.get("seed")) for x in (man.get("results") or [])
if x.get("seed") is not None) or "—"),
"- 建议:先听 listen/ 前奏与副歌片段;FAIL 则改词或换 seed 重 gen",
"",
])
path = os.path.join(rundir, "report.md")
with open(path, "w", encoding="utf-8") as f:
f.write(body)
return path
@bus.on("after_gen")
def _on_gen(run_dir=None, manifest=None, **_kw):
if not run_dir or not manifest:
return None
path = _write_report(run_dir, manifest)
return dict(report=path) if path else None
@bus.on("after_qc")
def _on_qc(run_dir=None, report=None, **_kw):
if not run_dir:
return None
man = {}
mp = os.path.join(run_dir, "manifest.json")
if os.path.isfile(mp):
with open(mp, encoding="utf-8") as f:
man = json.load(f)
intro = None
ip = os.path.join(run_dir, "intro_check.json")
if os.path.isfile(ip):
with open(ip, encoding="utf-8") as f:
intro = json.load(f)
path = _write_report(run_dir, man, qc=report, intro=intro)
return dict(report=path) if path else None
+3 -1
View File
@@ -5,7 +5,9 @@
before_build_tags kwargs: args, cfg, style, chill, instrumental, lean before_build_tags kwargs: args, cfg, style, chill, instrumental, lean
after_assemble kwargs: tags, lyrics, args, cfg, style, lean after_assemble kwargs: tags, lyrics, args, cfg, style, lean
-> 可返回 {"tags":..., "lyrics":...} 覆盖 -> 可返回 {"tags":..., "lyrics":...} 覆盖
before_qc_judge kwargs: result, expect, soft -> 可追加 why 列表 after_gen kwargs: run_dir, manifest, cfg, args
after_qc kwargs: run_dir, report, verd, rows, cfg, args
before_qc_judge kwargs: result, expect, soft (qc_worker 子进程内,可选)
after_qc_judge kwargs: verdict, why, result after_qc_judge kwargs: verdict, why, result
插件目录: 插件目录:
+1
View File
@@ -0,0 +1 @@
# package plugin
+60
View File
@@ -0,0 +1,60 @@
# -*- coding: utf-8 -*-
"""PASS 后截取前奏 + 副歌附近听感片段到 run/listen/。"""
import json
import os
from ..hooks import bus
def _clip(cfg, src, start, dur, out):
from ...cli.media import _run
code, blob = _run(dict(
op="clip", file=os.path.abspath(src),
start=float(start), dur=float(dur), out=os.path.abspath(out)))
return code == 0 and blob and not blob.get("error")
def _audio_from(cfg, man, rundir):
from ...cli.extras import _first_audio
return _first_audio(cfg, man or {}, rundir)
@bus.on("after_qc")
def _pack(run_dir=None, report=None, cfg=None, **_kw):
if not run_dir or not report or not cfg:
return None
verd = report.get("verdicts") or []
rows = report.get("rows") or []
new_names = {r.get("file") for r in rows if r.get("group") != "old"}
new_v = [v for v in verd if v.get("file") in new_names] or [
v for v in verd if v.get("verdict")]
if not new_v or any(v.get("verdict") != "PASS" for v in new_v):
return None
man = {}
mp = os.path.join(run_dir, "manifest.json")
if os.path.isfile(mp):
with open(mp, encoding="utf-8") as f:
man = json.load(f)
src = _audio_from(cfg, man, run_dir)
if not src or not os.path.isfile(src):
return None
bpm = float(man.get("bpm") or 104)
bars = int(man.get("intro_bars") or 8)
intro_sec = bars * 60.0 / bpm * 4.0
listen = os.path.join(run_dir, "listen")
os.makedirs(listen, exist_ok=True)
outs = []
a = os.path.join(listen, "intro.wav")
if _clip(cfg, src, 0.0, intro_sec, a):
outs.append(a)
# 副歌附近:约在 40% 时长处截 15s
dur = float(man.get("duration") or 180)
chorus_t = max(intro_sec + 1.0, dur * 0.40)
b = os.path.join(listen, "chorus_approx.wav")
if _clip(cfg, src, chorus_t, 15.0, b):
outs.append(b)
meta = dict(source=src, clips=outs, intro_sec=intro_sec,
chorus_start=chorus_t)
with open(os.path.join(listen, "pack.json"), "w", encoding="utf-8") as f:
json.dump(meta, f, ensure_ascii=False, indent=1)
return dict(listen=listen, clips=outs)
+1
View File
@@ -0,0 +1 @@
# package plugin
+54
View File
@@ -0,0 +1,54 @@
# -*- coding: utf-8 -*-
"""开曲衬词 FAIL 时给出重试建议;供 gen --auto-retry 读取。"""
import json
import os
from ..hooks import bus
# 判定文案关键字(与 qc_worker judge 一致)
_INTRO_KEYS = ("开曲衬词", "前 2s 就已唱出", "人声在")
def is_intro_fail(why_list):
text = " ".join(why_list or [])
return any(k in text for k in _INTRO_KEYS)
@bus.on("after_qc")
def _hint(run_dir=None, report=None, **_kw):
if not run_dir or not report:
return None
verd = report.get("verdicts") or []
reasons = []
for v in verd:
if v.get("verdict") == "FAIL":
reasons.extend(v.get("why") or [])
if not is_intro_fail(reasons):
# 也看 intro_check.json
ip = os.path.join(run_dir, "intro_check.json")
if os.path.isfile(ip):
with open(ip, encoding="utf-8") as f:
blob = json.load(f)
if blob.get("fail") or blob.get("ok") is False:
reasons.append("intro_check fail")
else:
return None
else:
return None
hint = dict(
retry=True,
reason="intro_filler_or_early_vocal",
why=reasons[:8],
suggest="换 seed 重 gen,或 songforge recheck --run <id> 后再决定",
)
path = os.path.join(run_dir, "retry_hint.json")
with open(path, "w", encoding="utf-8") as f:
json.dump(hint, f, ensure_ascii=False, indent=1)
return hint
def should_retry_intro_blob(blob):
"""给 gen --auto-retry 用:intro_check blob 是否应换 seed。"""
if not blob or blob.get("error"):
return False
return bool(blob.get("fail") or blob.get("ok") is False)
+65
View File
@@ -0,0 +1,65 @@
# -*- coding: utf-8 -*-
"""开曲轻检:只切前奏窗做 ASR 衬词统计(比全曲 QC 快)。
用法(qc_python):
python intro_check_worker.py '{"file":"...","bpm":104,"intro_bars":8,"lang":"zh"}'
"""
import json
import os
import sys
import traceback
HERE = os.path.dirname(os.path.abspath(__file__))
if HERE not in sys.path:
sys.path.insert(0, HERE)
def main(job):
import numpy as np
import qc_worker as q
path = job["file"]
bpm = float(job.get("bpm") or 104)
bars = int(job.get("intro_bars") or 8)
lang = job.get("lang") or "zh"
intro_sec = bars * 60.0 / bpm * 4.0
# 多取 0.5s 余量
take = min(intro_sec + 0.5, 60.0)
y, sr, ch = q._load_audio(path)
n = int(take * sr)
head = y[:max(1, n)]
a = q.asr(head, sr, lang)
segs = a.get("segments") or []
# 时间戳相对 clip 起点,整段都在前奏窗内
stats = q.intro_filler_stats(segs, intro_sec + 1.0, lang)
hard = q.HARD
fail = (stats["intro_filler_n"] >= hard["intro_filler_min_chars"]
and stats["intro_filler_ratio"] >= hard["intro_filler_ratio"])
# 前 2s 真唱字(content words)也算开曲违规
early = 0
for s in segs:
if float(s.get("t0", 0)) >= 2.0:
continue
early += len(q.content_words(s.get("text") or "", lang))
if early > 0:
fail = True
return dict(
ok=not fail,
fail=fail,
intro_sec=round(intro_sec, 2),
early_content=early,
asr_head=(a.get("text") or "")[:160],
**stats,
)
if __name__ == "__main__":
try:
job = json.loads(sys.argv[1])
out = main(job)
except Exception as e:
out = dict(error=str(e), type=type(e).__name__,
trace=traceback.format_exc()[-800:])
print("###JSON###")
print(json.dumps(out, ensure_ascii=False))