From 4ace5a95b052d338947fc88809a440ccd55cf6b4 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 16 十一月 2023 16:39:52 +0800
Subject: [PATCH] funasr pages

---
 funasr/utils/timestamp_tools.py |  102 +++++++++++++++++++++++++++++++++------------------
 1 files changed, 66 insertions(+), 36 deletions(-)

diff --git a/funasr/utils/timestamp_tools.py b/funasr/utils/timestamp_tools.py
index 09c3bec..6594273 100644
--- a/funasr/utils/timestamp_tools.py
+++ b/funasr/utils/timestamp_tools.py
@@ -1,14 +1,28 @@
-from pydoc import TextRepr
-from scipy.fftpack import shift
 import torch
-import copy
 import codecs
 import logging
-import edit_distance
 import argparse
-import pdb
 import numpy as np
-from typing import Any, List, Tuple, Union
+import edit_distance
+from itertools import zip_longest
+
+
+def cif_wo_hidden(alphas, threshold):
+    batch_size, len_time = alphas.size()
+    # loop varss
+    integrate = torch.zeros([batch_size], device=alphas.device)
+    # intermediate vars along time
+    list_fires = []
+    for t in range(len_time):
+        alpha = alphas[:, t]
+        integrate += alpha
+        list_fires.append(integrate)
+        fire_place = integrate >= threshold
+        integrate = torch.where(fire_place,
+                                integrate - torch.ones([batch_size], device=alphas.device)*threshold,
+                                integrate)
+    fires = torch.stack(list_fires, 1)
+    return fires
 
 
 def ts_prediction_lfr6_standard(us_alphas, 
@@ -19,24 +33,29 @@
                        sil_in_str=True
                        ):
     if not len(char_list):
-        return []
+        return "", []
     START_END_THRESHOLD = 5
     MAX_TOKEN_DURATION = 12
     TIME_RATE = 10.0 * 6 / 1000 / 3  #  3 times upsampled
     if len(us_alphas.shape) == 2:
-        _, peaks = us_alphas[0], us_peaks[0]  # support inference batch_size=1 only
+        alphas, peaks = us_alphas[0], us_peaks[0]  # support inference batch_size=1 only
     else:
-        _, peaks = us_alphas, us_peaks
-    num_frames = peaks.shape[0]
+        alphas, peaks = us_alphas, us_peaks
     if char_list[-1] == '</s>':
         char_list = char_list[:-1]
+    fire_place = torch.where(peaks>1.0-1e-4)[0].cpu().numpy() + force_time_shift  # total offset
+    if len(fire_place) != len(char_list) + 1:
+        alphas /= (alphas.sum() / (len(char_list) + 1))
+        alphas = alphas.unsqueeze(0)
+        peaks = cif_wo_hidden(alphas, threshold=1.0-1e-4)[0]
+        fire_place = torch.where(peaks>1.0-1e-4)[0].cpu().numpy() + force_time_shift  # total offset
+    num_frames = peaks.shape[0]
     timestamp_list = []
     new_char_list = []
     # for bicif model trained with large data, cif2 actually fires when a character starts
     # so treat the frames between two peaks as the duration of the former token
     fire_place = torch.where(peaks>1.0-1e-4)[0].cpu().numpy() + force_time_shift  # total offset
-    num_peak = len(fire_place)
-    assert num_peak == len(char_list) + 1 # number of peaks is supposed to be number of tokens + 1
+    # assert num_peak == len(char_list) + 1 # number of peaks is supposed to be number of tokens + 1
     # begin silence
     if fire_place[0] > START_END_THRESHOLD:
         # char_list.insert(0, '<sil>')
@@ -80,6 +99,7 @@
 
 
 def time_stamp_sentence(punc_id_list, time_stamp_postprocessed, text_postprocessed):
+    punc_list = ['锛�', '銆�', '锛�', '銆�']
     res = []
     if text_postprocessed is None:
         return res
@@ -89,44 +109,54 @@
         return res
     if len(text_postprocessed) == 0:
         return res
+
     if punc_id_list is None or len(punc_id_list) == 0:
         res.append({
             'text': text_postprocessed.split(),
             "start": time_stamp_postprocessed[0][0],
-            "end": time_stamp_postprocessed[-1][1]
+            "end": time_stamp_postprocessed[-1][1],
+            'text_seg': text_postprocessed.split(),
+            "ts_list": time_stamp_postprocessed,
         })
         return res
     if len(punc_id_list) != len(time_stamp_postprocessed):
-        res.append({
-            'text': text_postprocessed.split(),
-            "start": time_stamp_postprocessed[0][0],
-            "end": time_stamp_postprocessed[-1][1]
-        })
-        return res
-
-    sentence_text = ''
+        print("  warning length mistach!!!!!!")
+    sentence_text = ""
+    sentence_text_seg = ""
+    ts_list = []
     sentence_start = time_stamp_postprocessed[0][0]
+    sentence_end = time_stamp_postprocessed[0][1]
     texts = text_postprocessed.split()
-    for i in range(len(punc_id_list)):
-        sentence_text += texts[i]
-        if punc_id_list[i] == 2:
-            sentence_text += ','
+    punc_stamp_text_list = list(zip_longest(punc_id_list, time_stamp_postprocessed, texts, fillvalue=None))
+    for punc_stamp_text in punc_stamp_text_list:
+        punc_id, time_stamp, text = punc_stamp_text
+        # sentence_text += text if text is not None else ''
+        if text is not None:
+            if 'a' <= text[0] <= 'z' or 'A' <= text[0] <= 'Z':
+                sentence_text += ' ' + text
+            elif len(sentence_text) and ('a' <= sentence_text[-1] <= 'z' or 'A' <= sentence_text[-1] <= 'Z'):
+                sentence_text += ' ' + text
+            else:
+                sentence_text += text
+            sentence_text_seg += text + ' '
+        ts_list.append(time_stamp)
+
+        punc_id = int(punc_id) if punc_id is not None else 1
+        sentence_end = time_stamp[1] if time_stamp is not None else sentence_end
+
+        if punc_id > 1:
+            sentence_text += punc_list[punc_id - 2]
             res.append({
                 'text': sentence_text,
                 "start": sentence_start,
-                "end": time_stamp_postprocessed[i][1]
+                "end": sentence_end,
+                "text_seg": sentence_text_seg,
+                "ts_list": ts_list
             })
             sentence_text = ''
-            sentence_start = time_stamp_postprocessed[i][1]
-        elif punc_id_list[i] == 3:
-            sentence_text += '.'
-            res.append({
-                'text': sentence_text,
-                "start": sentence_start,
-                "end": time_stamp_postprocessed[i][1]
-            })
-            sentence_text = ''
-            sentence_start = time_stamp_postprocessed[i][1]
+            sentence_text_seg = ''
+            ts_list = []
+            sentence_start = sentence_end
     return res
 
 

--
Gitblit v1.9.1