# 黏著語模型爆爛

Examples
---

Kaldi 的utils/prepare_lang.sh有提供position-dependent-phones的選項，可以擴增每個phoneme的model(use _B, _E, _S & _I)結果在這樣的model下GMM的pipeline結果竟是：

training data: formosa_all 70000up + DAAI 70000up + Decept 7000up +  utterances
%WER 95.84 [ 16466 / 17181, 308 ins, 9340 del, 6818 sub ] exp_PDP/mono/decode_test0/cer_7_0.0
%WER 78.77 [ 13533 / 17181, 733 ins, 6507 del, 6293 sub ] exp_PDP/tri1/decode_test0/cer_10_0.0
%WER 77.38 [ 13294 / 17181, 583 ins, 7049 del, 5662 sub ] exp_PDP/tri2/decode_test0/cer_12_0.0
%WER 70.47 [ 12107 / 17181, 873 ins, 4877 del, 6357 sub ] exp_PDP/tri4a/decode_test0/cer_11_0.5
%WER 76.80 [ 13195 / 17181, 771 ins, 5862 del, 6562 sub ] exp_PDP/tri4a/decode_test0.si/cer_11_0.0
%WER 70.71 [ 12149 / 17181, 879 ins, 5072 del, 6198 sub ] exp_PDP/tri5a/decode_test0/cer_13_0.0
%WER 76.77 [ 13189 / 17181, 760 ins, 5651 del, 6778 sub ] exp_PDP/tri5a/decode_test0.si/cer_11_0.0

與用原本phoneme set的結果相比：

training data: formosa_small 18700 *2 (ADOSnoise)    utterances

%WER 82.41 [ 14159 / 17181, 453 ins, 6166 del, 7540 sub ] exp_mfcc/mono_noise0/decode_test0/cer_11_0.0
%WER 63.70 [ 10945 / 17181, 674 ins, 5112 del, 5159 sub ] exp_mfcc/tri1_noise0/decode_test0/cer_12_0.0
%WER 63.32 [ 10879 / 17181, 624 ins, 5126 del, 5129 sub ] exp_mfcc/tri2_noise0/decode_test0/cer_12_0.0
%WER 64.38 [ 11061 / 17181, 644 ins, 5302 del, 5115 sub ] exp_mfcc/tri3a_noise0/decode_test0/cer_16_0.0
%WER 55.35 [ 9509 / 17181, 758 ins, 4377 del, 4374 sub ] exp_mfcc/tri4a_noise0/decode_test0/cer_15_0.0
%WER 55.72 [ 9573 / 17181, 884 ins, 4250 del, 4439 sub ] exp_mfcc/tri5a_noise0/decode_test0/cer_14_0.0

大概差了20% 準確率。 或許是因為formosa 給的lexicon裡面就包含每個phoneme的四種抑揚頓挫發音，因此就沒必要設計position-dependent-phones了

position-dependent-phones
---

position-dependent-phones 的討論在2017 年Interspeech上有:[Improved subword modeling for WFST-based speech recognition](https://www.isca-speech.org/archive/Interspeech_2017/abstracts/0103.html)。

Themes
---
- [Dark theme](/theme-dark?both)
- [Vertical alignment](/theme-vertical-writing?both)

###### tags: `Kaldi`