冒险解谜游戏中文网 ChinaAVG
标题:
【游戏视频汉化 #1】 AI语音识别软件 Whisper编译
[打印本页]
作者:
shane007
时间:
2023-8-28 11:48
标题:
【游戏视频汉化 #1】 AI语音识别软件 Whisper编译
多年之前,就有为不带字幕的游戏视频配上字幕的想法。
$ n I1 e Y# N1 K7 G' y" G( }4 _
但是当时条件不成熟,但是目前来看,条件似乎成熟了
1 ]% A) b8 g" ?1 A/ o: @! _! `; \
: a0 V; R) F0 T0 Y H' Z
Whisper是openAI的开源语音识别软件。
4 h5 F7 I! `3 z% K6 x& M
它有一个.net的版本,在这个版本的基础上进行少量修改,就能将游戏视频对应的字幕识别成srt格式。
1 r% c0 Y3 U2 J: s, K2 H
之后,对这个srt文件再进行在线批量翻译之后,进行少量调整之后,汉化工作就完成了。
5 o# H, b3 r* d% }9 ?* v
8 {0 Z% h' o% T" w* J1 ]/ }
地址如下
! K# R3 F# }, o0 n
https://github.com/sandrohanea/whisper.net
9 d, v& ?$ y' y$ W
& s. @5 @( K' q' O; D5 e% Y# A+ V
4 }2 F/ \2 @" C9 T8 n3 h9 o; n3 I* i
编译最好使用vs2022编译,否则在.net sdk版本上会出很多问题。
d+ x q; |# X, c+ P+ i
o1 p( ?+ `; ]" m
编译好之后,有几个注意点
3 t' N( A3 o( M+ f2 K) I
6 n; U) C5 D6 P- l
<0>使用的模型文件修改为大模型,ggml-large.bin,用这个模型效果比较好。
8 {$ V. [4 s) N; J# j6 s
当然,所有时间也会比较多,估计转换一批文件需要几个甚至几十个小时。
! u3 T8 y+ W! J) X
, i& s1 X" |0 s$ v: ~
<1>Language要设定为"english"。
3 M# e( P: n! N1 m( S9 `5 J
* H/ d$ Y; `8 i6 M/ Z
/* var builder = factory.CreateBuilder()
# a0 `7 A9 C$ @1 ]
.WithLanguage(opt.Language);*/
& k; d0 x; T5 L# R! m u
var builder = factory.CreateBuilder()
; |0 D3 |# F( x
.WithLanguage("english");
复制代码
1 A( ]+ W* w; N1 k# X
<2>缺省好像只支持Wav格式,而且是要16K采样率的,需要实现转换成这种格式,否则会出错。
4 P Q! f; Q1 e6 a# P
* }" o1 _$ H4 s/ m8 v. r, S
<3>缺省只提供了一个例子wav文件的转换,需要改为批量形式。
! p# H6 Y9 q; q7 ~+ t+ V
(遍历某个目录中的所有文件)
8 Z- O+ c( O4 g6 X3 |% e, O( n
( b( R% g9 o8 P- q/ A5 e! f9 c; T) z
<4>输出的文件,需要稍加整理,以符合srt格式
/ q5 A! h' J$ x; H# h' B- T! C
3 C; O2 X+ L% M( ?: S4 S7 z
以下是一个Wav文件的控制台输出(幽魂开场动画)
( L% p6 a* }) f5 p3 W3 T& k
4 W. [/ z3 }1 r2 A* r% ]
# d$ p9 B: Z' `" j F5 R6 x5 G0 Z v
whisper_init_from_file_no_state: loading model from 'ggml-large.bin'
& _8 b& T# w! t- y) N9 C o4 I3 c
whisper_model_load: loading model
! ], G( ?# k* G! f% F4 i/ v3 ?
whisper_model_load: n_vocab = 51865
4 y! W$ ~! R# |
whisper_model_load: n_audio_ctx = 1500
5 Y- X$ d( z0 i
whisper_model_load: n_audio_state = 1280
3 J3 s& I. R( V& G
whisper_model_load: n_audio_head = 20
5 P9 N" O! Z9 d$ `6 ?4 Z7 g
whisper_model_load: n_audio_layer = 32
$ ~& N: D1 f: Q5 O
whisper_model_load: n_text_ctx = 448
7 w1 {, ~* a& Z! P" x0 f7 y8 M$ F
whisper_model_load: n_text_state = 1280
( `8 |& e: B: W8 `2 d' k1 }. m
whisper_model_load: n_text_head = 20
9 ^. i; S. p8 J3 m) W* v
whisper_model_load: n_text_layer = 32
7 Y: i0 a8 n8 ?5 G7 Y& x' l9 e! g
whisper_model_load: n_mels = 80
1 I1 H2 l' v. D5 Y& c
whisper_model_load: ftype = 1
7 m( K9 n2 S4 b9 o: x4 ^& k% F; o
whisper_model_load: qntvr = 0
6 r* V% v! {6 \3 f; o' h
whisper_model_load: type = 5
8 R8 J' M* e( B
whisper_model_load: mem required = 3557.00 MB (+ 71.00 MB per decoder)
- B+ }9 Y) l; }; O6 P
whisper_model_load: adding 1608 extra tokens
/ [% k9 u+ S1 p2 L/ L2 c
whisper_model_load: model ctx = 2951.27 MB
, T7 \- Z! k& N: D" g
whisper_model_load: model size = 2950.66 MB
0 X5 U) K/ p" K# D% b9 S
whisper_init_state: kv self size = 70.00 MB
" F4 p* e/ J/ G Z4 H
whisper_init_state: kv cross size = 234.38 MB
) L B! c3 k" Z; P) O+ ~+ I! D
New Segment: 00:00:00 ==> 00:00:02.7600000 : (birds chirping)
4 v3 F k& \4 R7 ?
New Segment: 00:00:03.6600000 ==> 00:00:05.9000000 : (exhaling)
/ p- f8 z1 y9 {* ~, ~2 ~
New Segment: 00:00:05.9000000 ==> 00:00:08.6600000 : (birds chirping)
( A3 J& B8 x+ Y( |! w( W
New Segment: 00:00:08.6600000 ==> 00:00:35.1200000 : (gun firing)
1 \1 |- u$ A Q+ x+ f
New Segment: 00:00:36.1200000 ==> 00:00:38.5400000 : (gun firing)
0 F7 Y5 Q3 T, H! p
New Segment: 00:00:39.0600000 ==> 00:00:41.4800000 : (gun firing)
: L4 z0 p0 o1 \ B1 ?: _4 J' I
New Segment: 00:00:41.4800000 ==> 00:00:49.4000000 : (tires screeching)
5 _- U6 c$ \' \! ~) `- D: }& A* @
New Segment: 00:00:49.4000000 ==> 00:00:58.5800000 : (glass shattering)
7 o: i$ H3 t' [" p O! j
New Segment: 00:00:58.5800000 ==> 00:01:07.7400000 : (singing in foreign language)
, \, _5 A K/ W' l% `. F
New Segment: 00:01:07.7400000 ==> 00:01:11.5800000 : (singing in foreign language)
* e9 R3 Z6 ~6 l: g
New Segment: 00:01:11.5800000 ==> 00:01:17 : (tires screeching)
9 A( \' q2 D' R' X
New Segment: 00:01:17 ==> 00:01:24.8400000 : (singing in foreign language)
) X8 l. [- a+ }" F8 u/ m
New Segment: 00:01:24.8400000 ==> 00:01:28.6400000 : (panting)
) A+ R5 n5 N6 n* J0 z
New Segment: 00:01:36.7800000 ==> 00:01:39.2000000 : (gun firing)
, \+ ]! X& y8 S
New Segment: 00:01:39.2000000 ==> 00:01:43.4600000 : - Adrian.
1 T' A2 T3 [/ l1 [+ z) Y {. r
New Segment: 00:01:43.4600000 ==> 00:01:45.6200000 : - Oh God.
1 K* E8 l" f, I) O: `) ?# P* p7 m
New Segment: 00:01:45.6200000 ==> 00:01:48.2000000 : - What's the matter sweetheart?
' |, Q6 [) ]% ]+ x- f
New Segment: 00:01:48.2000000 ==> 00:01:50.4200000 : Oh.
/ ~" L. e) Z5 M8 ?. x
New Segment: 00:01:50.4200000 ==> 00:01:53.4600000 : - Oh it's horrible.
8 c+ u' V' q. `0 p2 C
New Segment: 00:01:53.4600000 ==> 00:01:55.3000000 : - Shh.
( _. L7 r3 T3 L2 z) ]/ F
New Segment: 00:01:55.3000000 ==> 00:02:02.3400000 : It was just a bad dream.
; p# R2 q+ |% d
New Segment: 00:02:05.4200000 ==> 00:02:09.8800000 : - You don't ever have to be afraid of anything.
, ?; J- R' \& B; G4 C
New Segment: 00:02:09.8800000 ==> 00:02:12.8000000 : I'll always be here to protect you.
0 e' W$ ~$ u5 n. O4 V* K
New Segment: 00:02:12.9200000 ==> 00:02:15.5000000 : (gentle music)
0 I0 }, ]) r, l; b1 f5 ^! [
New Segment: 00:02:16.4800000 ==> 00:02:19.0600000 : (gentle music)
5 R2 t# y4 c6 A( g" ]: ~; y
New Segment: 00:02:19.0600000 ==> 00:02:21.6400000 : (gentle music)
) E) e' }. ?7 R8 M9 u$ ^5 Y
New Segment: 00:02:21.6400000 ==> 00:02:24.2200000 : (gentle music)
$ C: U2 W- j$ U
New Segment: 00:02:24.5400000 ==> 00:02:27.1200000 : (gentle music)
& n6 w, O+ @6 u7 R( {
New Segment: 00:02:27.1200000 ==> 00:02:29.7000000 : (gentle music)
! i6 Q+ `& r. K" k! l
New Segment: 00:02:29.7000000 ==> 00:02:33.1800000 : [Music]
% }- [- N4 I- e, R" k; m
复制代码
" } T3 d* ~2 s" S$ j& g2 n
/ O, c4 C' C$ X; ?, V0 }* O
欢迎光临 冒险解谜游戏中文网 ChinaAVG (https://chinaavg.com/)
Powered by Discuz! X3.2