update timestamp and add ploter

funasr/runtime/python/onnxruntime/demo.py

@@ -6,7 +6,7 @@ model_dir = "/Users/shixian/code/funasr2/export/damo/speech_paraformer-large-vad
 
 model = Paraformer(model_dir, batch_size=1)
 
-wav_path = ['/Users/shixian/code/funasr2/export/damo/speech_paraformer-tiny-commandword_asr_nat-zh-cn-16k-vocab544-pytorch/example/asr_example.wav']
+wav_path = ['/Users/shixian/code/funasr2/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/example/asr_example.wav']
 
 result = model(wav_path)
 print(result)

funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py

@@ -1,6 +1,7 @@
 # -*- encoding: utf-8 -*-
 # @Author: SWHL
 # @Contact: liekkaskono@163.com
+from cgitb import text
 import os.path
 from pathlib import Path
 from typing import List, Union, Tuple
@@ -23,6 +24,7 @@ class Paraformer():
     def __init__(self, model_dir: Union[str, Path] = None,
                  batch_size: int = 1,
                  device_id: Union[str, int] = "-1",
+                 plot_timestamp: bool = True,
                  ):
 
         if not Path(model_dir).exists():
@@ -41,7 +43,7 @@ class Paraformer():
         )
         self.ort_infer = OrtInferSession(model_file, device_id)
         self.batch_size = batch_size
-        self.plot = True
+        self.plot = plot_timestamp
 
     def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs) -> List:
         waveform_list = self.load_data(wav_content, self.frontend.opts.frame_opts.samp_freq)
@@ -76,7 +78,26 @@ class Paraformer():
 
     def plot_wave_timestamp(self, wav, text_timestamp):
         # TODO: Plot the wav and timestamp results with matplotlib
-        import pdb; pdb.set_trace()
+        import matplotlib
+        matplotlib.use('Agg')
+        matplotlib.rc("font", family='Alibaba PuHuiTi')  # set it to a font that your system supports
+        import matplotlib.pyplot as plt
+        fig, ax1 = plt.subplots(figsize=(11, 3.5), dpi=320)
+        ax2 = ax1.twinx()
+        ax2.set_ylim([0, 2.0])
+        # plot waveform
+        ax1.set_ylim([-0.3, 0.3])
+        time = np.arange(wav.shape[0]) / 16000
+        ax1.plot(time, wav/wav.max()*0.3, color='gray', alpha=0.4)
+        # plot lines and text
+        for (char, start, end) in text_timestamp:
+            ax1.vlines(start, -0.3, 0.3, ls='--')
+            ax1.vlines(end, -0.3, 0.3, ls='--')
+            x_adj = 0.045 if char != '<sil>' else 0.12
+            ax1.text((start + end) * 0.5 - x_adj, 0, char)
+        # plt.legend()
+        plotname = "debug.png"
+        plt.savefig(plotname, bbox_inches='tight')
 
     def load_data(self,
                   wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List:

funasr/runtime/python/onnxruntime/rapid_paraformer/utils/timestamp_utils.py

@@ -1,11 +1,11 @@
 import numpy as np
 
 
-def time_stamp_lfr6_onnx(us_cif_peak, char_list, begin_time=0.0):
+def time_stamp_lfr6_onnx(us_cif_peak, char_list, begin_time=0.0, total_offset=-1.5):
     if not len(char_list):
         return []
     START_END_THRESHOLD = 5
-    MAX_TOKEN_DURATION = 14
+    MAX_TOKEN_DURATION = 30
     TIME_RATE = 10.0 * 6 / 1000 / 3  #  3 times upsampled
     cif_peak = us_cif_peak.reshape(-1)
     num_frames = cif_peak.shape[-1]
@@ -16,7 +16,7 @@ def time_stamp_lfr6_onnx(us_cif_peak, char_list, begin_time=0.0):
     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 = np.where(cif_peak>1.0-1e-4)[0] - 1.5  # np format
+    fire_place = np.where(cif_peak>1.0-1e-4)[0] + total_offset  # np format
     num_peak = len(fire_place)
     assert num_peak == len(char_list) + 1 # number of peaks is supposed to be number of tokens + 1
     # begin silence
@@ -27,7 +27,7 @@ def time_stamp_lfr6_onnx(us_cif_peak, char_list, begin_time=0.0):
     # tokens timestamp
     for i in range(len(fire_place)-1):
         new_char_list.append(char_list[i])
-        if MAX_TOKEN_DURATION < 0 or fire_place[i+1] - fire_place[i] < MAX_TOKEN_DURATION:
+        if i == len(fire_place)-2 or MAX_TOKEN_DURATION < 0 or fire_place[i+1] - fire_place[i] < MAX_TOKEN_DURATION:
             timestamp_list.append([fire_place[i]*TIME_RATE, fire_place[i+1]*TIME_RATE])
         else:
             # cut the duration to token and sil of the 0-weight frames last long
@@ -55,4 +55,5 @@ def time_stamp_lfr6_onnx(us_cif_peak, char_list, begin_time=0.0):
     for char, timestamp in zip(new_char_list, timestamp_list):
         if char != '<sil>':
             res.append([int(timestamp[0] * 1000), int(timestamp[1] * 1000)])
-    return res, res_total
+    return res, res_total
+