HY-2012 commited on
Commit
cf28c2e
·
verified ·
1 Parent(s): f1a72bd

CuteTTS.AXERA: AX650 NPU zero-shot TTS (U16)

Browse files
README.md CHANGED
@@ -108,9 +108,12 @@ bash run_ax650.sh "今天天气很好,我们一起去公园散步吧。" sampl
108
 
109
  # 调 CFG / 随机种子 / 最长 patch 数(1 patch ≈ 160 ms)
110
  bash run_ax650.sh "Hello from AX650!" samples/reference.wav en.wav --cfg 2.0 --seed 42 --max-new 128
 
 
 
111
  ```
112
 
113
- 参考音频 ≤ 5 s;单次生成 ≤ 20.5 s;输出 24 kHz 单声道。
114
 
115
  <a id="results"></a>
116
 
 
108
 
109
  # 调 CFG / 随机种子 / 最长 patch 数(1 patch ≈ 160 ms)
110
  bash run_ax650.sh "Hello from AX650!" samples/reference.wav en.wav --cfg 2.0 --seed 42 --max-new 128
111
+
112
+ # 长文本:自动按句切分(每段 ≤18 s),逐段生成后拼接
113
+ bash run_ax650.sh "长文本内容……" samples/reference.wav long.wav --split
114
  ```
115
 
116
+ 参考音频 ≤ 5 s;单次生成 ≤ 20.5 s(**超过加 `--split` 自动按句切分**,可支持任意长度);输出 24 kHz 单声道。
117
 
118
  <a id="results"></a>
119
 
bin/cutetts_ax650 CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:af84a6c7ea26f1e49b929cd968ba5b7f614b48d289269ff9e271c524e15bfd6e
3
- size 101240
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:887b6d2eed0453d943e404156c37b87df7eb8063b6e4932f71ddafaf8a0f86ab
3
+ size 125848
python/cutetts_ax/__init__.py CHANGED
@@ -1 +1,2 @@
1
  from .tts import CuteTTSAX, AxEngine, OnnxEngine # noqa: F401
 
 
1
  from .tts import CuteTTSAX, AxEngine, OnnxEngine # noqa: F401
2
+ from .text import estimate_seconds, split_text # noqa: E402,F401
python/cutetts_ax/text.py ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """长文本按句切分(单次请求上限约 20.3 s 语音)。
2
+
3
+ 规则与 C++ 实现(`cpp/src/text_split.hpp`)一致:
4
+ 1. 句末标点切分:。!?;…! ? ; 与换行
5
+ 2. 单句仍超长 → 退到 ,、, 处切 → 再不行按字符硬切
6
+ 3. 贪心合并到 max_seconds(默认 18 s,给 20.3 s 硬上限留余量)
7
+ """
8
+
9
+ from __future__ import annotations
10
+
11
+ import re
12
+
13
+ _SENT_END = set("。!?;…!?;.\n")
14
+ _SUB_END = set(",、,")
15
+ # 句末标点后可能跟的收尾符号,并入前一句
16
+ _TRAILING = set("\"'”』 」))]】")
17
+
18
+
19
+ def char_seconds(ch: str) -> float:
20
+ """单个字符的估算口语时长(秒)。"""
21
+ o = ord(ch)
22
+ if 0x4E00 <= o <= 0x9FFF: # 汉字
23
+ return 0.22
24
+ if 0x3040 <= o <= 0x30FF: # 假名
25
+ return 0.20
26
+ if ch.isascii() and (ch.isalnum()):
27
+ return 0.055
28
+ if ch in " ,-":
29
+ return 0.02
30
+ return 0.0
31
+
32
+
33
+ def estimate_seconds(text: str) -> float:
34
+ return sum(char_seconds(c) for c in text)
35
+
36
+
37
+ def _split_keep(text: str, marks: set) -> list[str]:
38
+ """在指定标点后切分,标点留在前一段。"""
39
+ out, buf = [], []
40
+ i = 0
41
+ n = len(text)
42
+ while i < n:
43
+ ch = text[i]
44
+ buf.append(ch)
45
+ i += 1
46
+ if ch in marks:
47
+ while i < n and (text[i] in _TRAILING or text[i] in " \t"):
48
+ buf.append(text[i])
49
+ i += 1
50
+ seg = "".join(buf)
51
+ if estimate_seconds(seg) > 0:
52
+ out.append(seg)
53
+ buf = []
54
+ if buf:
55
+ seg = "".join(buf)
56
+ if estimate_seconds(seg) > 0:
57
+ out.append(seg)
58
+ return out
59
+
60
+
61
+ def _hard_split(text: str, max_seconds: float) -> list[str]:
62
+ out, buf, t = [], [], 0.0
63
+ for ch in text:
64
+ d = char_seconds(ch)
65
+ if t + d > max_seconds and buf:
66
+ out.append("".join(buf))
67
+ buf, t = [], 0.0
68
+ buf.append(ch)
69
+ t += d
70
+ if buf:
71
+ out.append("".join(buf))
72
+ return out
73
+
74
+
75
+ def split_text(text: str, max_seconds: float = 18.0) -> list[str]:
76
+ """把长文本切成若干段,每段估算时长 ≤ max_seconds(尽量按句边界)。"""
77
+ if max_seconds <= 0:
78
+ max_seconds = 18.0
79
+ hard = max_seconds * 1.02
80
+
81
+ pieces: list[str] = []
82
+ for sent in _split_keep(text, _SENT_END):
83
+ if estimate_seconds(sent) <= hard:
84
+ pieces.append(sent)
85
+ continue
86
+ for sub in _split_keep(sent, _SUB_END):
87
+ if estimate_seconds(sub) <= hard:
88
+ pieces.append(sub)
89
+ else:
90
+ pieces.extend(_hard_split(sub, max_seconds))
91
+ if not pieces:
92
+ return [text]
93
+
94
+ merged: list[str] = []
95
+ cur = ""
96
+ cur_t = 0.0
97
+ for p in pieces:
98
+ pt = estimate_seconds(p)
99
+ if not cur:
100
+ cur, cur_t = p, pt
101
+ elif cur_t + pt <= max_seconds:
102
+ cur += p
103
+ cur_t += pt
104
+ else:
105
+ merged.append(cur)
106
+ cur, cur_t = p, pt
107
+ if cur:
108
+ merged.append(cur)
109
+ return merged or [text]
python/cutetts_ax/tts.py CHANGED
@@ -375,6 +375,53 @@ class CuteTTSAX:
375
  return max(self._dec_paths)
376
 
377
  # --------------------------------------------------------------------- api
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
378
  def synthesize(self, text, reference_audio, max_new=None, cfg_strength=None, seed=42,
379
  ref_max_seconds=5.0, rng=None):
380
  t0 = time.time()
 
375
  return max(self._dec_paths)
376
 
377
  # --------------------------------------------------------------------- api
378
+ @staticmethod
379
+ def _merge_timing(agg, tm):
380
+ """多段合成时聚合计时:TTFT/TTFA 取第一段,其余累加。"""
381
+ if tm is None:
382
+ return agg
383
+ if agg is None:
384
+ return dict(tm)
385
+ for k in ("total", "audio_seconds", "steps", "chunks", "dit", "lm_decode"):
386
+ if k in tm:
387
+ agg[k] = agg.get(k, 0) + tm[k]
388
+ return agg
389
+
390
+ # ---------------------------------------------------------------- long text
391
+ def synthesize_long(self, text, reference_audio, max_seconds=18.0, **kw):
392
+ """长文本:按句切分后逐段合成,返回拼接好的 PCM。
393
+
394
+ 单次请求上限约 20.3 s(KV 256 prompt + 127 patch),超出会被截断;
395
+ 这里按 ``max_seconds``(默认 18 s)切段,逐段生成后拼起来。
396
+ """
397
+ from .text import split_text
398
+
399
+ parts = split_text(text, max_seconds)
400
+ self.segments = len(parts)
401
+ outs, agg = [], None
402
+ for p in parts:
403
+ outs.append(self.synthesize(p, reference_audio, **kw))
404
+ agg = self._merge_timing(agg, self.last_timing)
405
+ if agg is not None:
406
+ agg["segments"] = len(parts)
407
+ self.last_timing = agg
408
+ return np.concatenate(outs) if outs else np.zeros(0, np.float32)
409
+
410
+ def synthesize_stream_long(self, text, reference_audio, max_seconds=18.0, **kw):
411
+ """长文本流式:按句切段,一段接一段地流式输出(音色缓存后每段 ≈0.5 s 首包)。"""
412
+ from .text import split_text
413
+
414
+ parts = split_text(text, max_seconds)
415
+ self.segments = len(parts)
416
+ agg = None
417
+ for p in parts:
418
+ yield from self.synthesize_stream(p, reference_audio, **kw)
419
+ agg = self._merge_timing(agg, self.last_timing)
420
+ if agg is not None:
421
+ agg["segments"] = len(parts)
422
+ agg["streaming"] = True
423
+ self.last_timing = agg
424
+
425
  def synthesize(self, text, reference_audio, max_new=None, cfg_strength=None, seed=42,
426
  ref_max_seconds=5.0, rng=None):
427
  t0 = time.time()
python/demo.py CHANGED
@@ -30,6 +30,9 @@ def main():
30
  ap.add_argument("--stream", action="store_true", help="true streaming: emit audio while generating")
31
  ap.add_argument("--first-chunk", type=int, default=8, help="latents in the first streamed chunk")
32
  ap.add_argument("--chunk", type=int, default=16, help="latents per streamed chunk afterwards")
 
 
 
33
  ap.add_argument("--stateful", action="store_true",
34
  help="per-patch stateful decoder (lowest latency, slightly lower quality)")
35
  a = ap.parse_args()
@@ -57,8 +60,13 @@ def main():
57
  f"(RTF {tm['total']/max(tm['audio_seconds'],1e-6):.3f})")
58
  print("saved", a.output)
59
  return
60
- wav = tts.synthesize(a.text, a.reference, max_new=a.max_new, cfg_strength=a.cfg,
61
- seed=a.seed, ref_max_seconds=a.ref_seconds)
 
 
 
 
 
62
  A.write_wav(a.output, wav, tts.sr)
63
  tm = tts.last_timing
64
  print(f"audio {tm['audio_seconds']:.2f}s in {tm['total']:.2f}s "
 
30
  ap.add_argument("--stream", action="store_true", help="true streaming: emit audio while generating")
31
  ap.add_argument("--first-chunk", type=int, default=8, help="latents in the first streamed chunk")
32
  ap.add_argument("--chunk", type=int, default=16, help="latents per streamed chunk afterwards")
33
+ ap.add_argument("--split", action="store_true",
34
+ help="长文本按句切分(每段 ≤ --split-sec 秒),逐段生成后拼接")
35
+ ap.add_argument("--split-sec", type=float, default=18.0)
36
  ap.add_argument("--stateful", action="store_true",
37
  help="per-patch stateful decoder (lowest latency, slightly lower quality)")
38
  a = ap.parse_args()
 
60
  f"(RTF {tm['total']/max(tm['audio_seconds'],1e-6):.3f})")
61
  print("saved", a.output)
62
  return
63
+ if a.split:
64
+ wav = tts.synthesize_long(a.text, a.reference, max_seconds=a.split_sec,
65
+ max_new=a.max_new, cfg_strength=a.cfg, seed=a.seed,
66
+ ref_max_seconds=a.ref_seconds)
67
+ else:
68
+ wav = tts.synthesize(a.text, a.reference, max_new=a.max_new, cfg_strength=a.cfg,
69
+ seed=a.seed, ref_max_seconds=a.ref_seconds)
70
  A.write_wav(a.output, wav, tts.sr)
71
  tm = tts.last_timing
72
  print(f"audio {tm['audio_seconds']:.2f}s in {tm['total']:.2f}s "