Dev gzf exp (#1654)

zhifu gao

2024-04-24 861147c7308b91068ffa02724fdf74ee623a909e

 runtime/python/onnxruntime/funasr_onnx/utils/timestamp_utils.py

@@ -13,7 +13,7 @@
    TIME_RATE = 10.0 * 6 / 1000 / 3  #  3 times upsampled
    cif_peak = us_cif_peak.reshape(-1)
    num_frames = cif_peak.shape[-1]
    if char_list[-1] == '</s>':
    if char_list[-1] == "</s>":
        char_list = char_list[:-1]
    # char_list = [i for i in text]
    timestamp_list = []
@@ -27,18 +27,22 @@
    if fire_place[0] > START_END_THRESHOLD:
        # char_list.insert(0, '<sil>')
        timestamp_list.append([0.0, fire_place[0]*TIME_RATE])
        new_char_list.append('<sil>')
        new_char_list.append("<sil>")
    # tokens timestamp
    for i in range(len(fire_place)-1):
        new_char_list.append(char_list[i])
        if i == len(fire_place)-2 or 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
            _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('<sil>')
            new_char_list.append("<sil>")
    # tail token and end silence
    if num_frames - fire_place[-1] > START_END_THRESHOLD:
        _end = (num_frames + fire_place[-1]) / 2
@@ -57,7 +61,6 @@
        res_str += "{} {} {};".format(char, timestamp[0], timestamp[1])
    res = []
    for char, timestamp in zip(new_char_list, timestamp_list):
        if char != '<sil>':
        if char != "<sil>":
            res.append([int(timestamp[0] * 1000), int(timestamp[1] * 1000)])
    return res_str, res