mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-26 10:16:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

add alimeeting norm

Browse Source
This commit is contained in:
Yuekai Zhang 2024-06-26 10:41:49 +08:00
parent b594a3875b
commit 9a5fc2ab71
1 changed files with 39 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

39
egs/multi_zh-hans/ASR/zipformer/train.py
View File

@ -775,6 +775,42 @@ def compute_loss(
warmup: a floating point value which increases throughout training; warmup: a floating point value which increases throughout training;
values >= 1.0 are fully warmed up and have all modules present. values >= 1.0 are fully warmed up and have all modules present.
""" """
def normalize_text_alimeeting(text: str, normalize: str = "m2met") -> str:
"""
Text normalization similar to M2MeT challenge baseline.
See: https://github.com/yufan-aslp/AliMeeting/blob/main/asr/local/text_normalize.pl
"""
if normalize == "none":
return text
elif normalize == "m2met":
import re
text = text.replace(" ", "")
text = text.replace("<sil>", "")
text = text.replace("<%>", "")
text = text.replace("<->", "")
text = text.replace("<$>", "")
text = text.replace("<#>", "")
text = text.replace("<_>", "")
text = text.replace("<space>", "")
text = text.replace("`", "")
text = text.replace("&", "")
text = text.replace(",", "")
if re.search("[a-zA-Z]", text):
text = text.upper()
text = text.replace("Ａ", "A")
text = text.replace("ａ", "A")
text = text.replace("ｂ", "B")
text = text.replace("ｃ", "C")
text = text.replace("ｋ", "K")
text = text.replace("ｔ", "T")
text = text.replace("，", "")
text = text.replace("丶", "")
text = text.replace("。", "")
text = text.replace("、", "")
text = text.replace("？", "")
return text
device = model.device if isinstance(model, DDP) else next(model.parameters()).device device = model.device if isinstance(model, DDP) else next(model.parameters()).device
feature = batch["inputs"] feature = batch["inputs"]
# at entry, feature is (N, T, C) # at entry, feature is (N, T, C)
@ -788,6 +824,9 @@ def compute_loss(
warm_step = params.warm_step warm_step = params.warm_step
texts = batch["supervisions"]["text"] texts = batch["supervisions"]["text"]
# remove spaces in texts
texts = [normalize_text_alimeeting(text) for text in texts]
y = sp.encode(texts, out_type=int) y = sp.encode(texts, out_type=int)
y = k2.RaggedTensor(y) y = k2.RaggedTensor(y)