From 9dd4901aade665eb64f67cf8b58454c51fe1cf33 Mon Sep 17 00:00:00 2001 From: "shixian.shi" Date: Thu, 2 Mar 2023 19:50:06 +0800 Subject: [PATCH] rapid_paraformer.utils.timestamp_utils --- funasr/runtime/python/onnxruntime/demo.py | 1 + .../rapid_paraformer/paraformer_onnx.py | 4 +- .../rapid_paraformer/utils/timestamp_utils.py | 59 +++++++++++++++++++ funasr/utils/timestamp_tools.py | 1 + 4 files changed, 63 insertions(+), 2 deletions(-) create mode 100644 funasr/runtime/python/onnxruntime/rapid_paraformer/utils/timestamp_utils.py diff --git a/funasr/runtime/python/onnxruntime/demo.py b/funasr/runtime/python/onnxruntime/demo.py index 48ad6ed53..d35960383 100644 --- a/funasr/runtime/python/onnxruntime/demo.py +++ b/funasr/runtime/python/onnxruntime/demo.py @@ -3,6 +3,7 @@ from rapid_paraformer import Paraformer model_dir = "/Users/shixian/code/funasr2/export/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch" model_dir = "/Users/shixian/code/funasr2/export/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" + 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'] diff --git a/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py b/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py index 8af347444..ed9b030df 100644 --- a/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py +++ b/funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py @@ -14,7 +14,7 @@ from .utils.utils import (CharTokenizer, Hypothesis, ONNXRuntimeError, read_yaml) from .utils.postprocess_utils import sentence_postprocess from .utils.frontend import WavFrontend -from funasr.utils.timestamp_tools import time_stamp_lfr6_pl +from .utils.timestamp_utils import time_stamp_lfr6_onnx logging = get_logger() @@ -68,7 +68,7 @@ class Paraformer(): preds, raw_token = self.decode(am_scores, valid_token_lens)[0] res['preds'] = preds if us_cif_peak is not None: - timestamp = time_stamp_lfr6_pl(us_alphas, us_cif_peak, copy.copy(raw_token), log=False) + timestamp = time_stamp_lfr6_onnx(us_cif_peak, copy.copy(raw_token)) res['timestamp'] = timestamp asr_res.append(res) return asr_res diff --git a/funasr/runtime/python/onnxruntime/rapid_paraformer/utils/timestamp_utils.py b/funasr/runtime/python/onnxruntime/rapid_paraformer/utils/timestamp_utils.py new file mode 100644 index 000000000..32dd8498e --- /dev/null +++ b/funasr/runtime/python/onnxruntime/rapid_paraformer/utils/timestamp_utils.py @@ -0,0 +1,59 @@ +import numpy as np + + +def time_stamp_lfr6_onnx(us_cif_peak, char_list, begin_time=0.0): + if not len(char_list): + return [] + START_END_THRESHOLD = 5 + MAX_TOKEN_DURATION = 14 + TIME_RATE = 10.0 * 6 / 1000 / 3 # 3 times upsampled + cif_peak = us_cif_peak.reshape(-1) + num_frames = cif_peak.shape[-1] + import pdb; pdb.set_trace() + if char_list[-1] == '': + char_list = char_list[:-1] + # char_list = [i for i in text] + 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 = np.where(cif_peak>1.0-1e-4)[0] - 1.5 # 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 + if fire_place[0] > START_END_THRESHOLD: + # char_list.insert(0, '') + timestamp_list.append([0.0, fire_place[0]*TIME_RATE]) + new_char_list.append('') + # 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: + 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 + _split = fire_place[i] + MAX_TOKEN_DURATION + timestamp_list.append([fire_place[i]*TIME_RATE, _split*TIME_RATE]) + timestamp_list.append([_split*TIME_RATE, fire_place[i+1]*TIME_RATE]) + new_char_list.append('') + # tail token and end silence + if num_frames - fire_place[-1] > START_END_THRESHOLD: + _end = (num_frames + fire_place[-1]) / 2 + timestamp_list[-1][1] = _end*TIME_RATE + timestamp_list.append([_end*TIME_RATE, num_frames*TIME_RATE]) + new_char_list.append("") + else: + timestamp_list[-1][1] = num_frames*TIME_RATE + if begin_time: # add offset time in model with vad + for i in range(len(timestamp_list)): + timestamp_list[i][0] = timestamp_list[i][0] + begin_time / 1000.0 + timestamp_list[i][1] = timestamp_list[i][1] + begin_time / 1000.0 + assert len(new_char_list) == len(timestamp_list) + res_txt = "" + for char, timestamp in zip(new_char_list, timestamp_list): + res_txt += "{} {} {};".format(char, timestamp[0], timestamp[1]) + res = [] + for char, timestamp in zip(new_char_list, timestamp_list): + if char != '': + res.append([int(timestamp[0] * 1000), int(timestamp[1] * 1000)]) + return res \ No newline at end of file diff --git a/funasr/utils/timestamp_tools.py b/funasr/utils/timestamp_tools.py index b82c74ad7..4a367f8ec 100644 --- a/funasr/utils/timestamp_tools.py +++ b/funasr/utils/timestamp_tools.py @@ -55,6 +55,7 @@ def time_stamp_lfr6_pl(us_alphas, us_cif_peak, char_list, begin_time=0.0, end_ti res.append([int(timestamp[0] * 1000), int(timestamp[1] * 1000)]) return res + def time_stamp_sentence(punc_id_list, time_stamp_postprocessed, text_postprocessed): res = [] if text_postprocessed is None: