Bug in whisper word-level timestamps (`tokenizer._decode_asr`)

[xenova]

System Info

• transformers version: 4.42.3
• Platform: Linux-6.1.85+-x86_64-with-glibc2.35
• Python version: 3.10.12
• Huggingface_hub version: 0.23.4
• Safetensors version: 0.4.3
• Accelerate version: not installed
• Accelerate config: not found
• PyTorch version (GPU?): 2.3.0+cu121 (False)
• Tensorflow version (GPU?): 2.15.0 (False)
• Flax version (CPU?/GPU?/TPU?): 0.8.4 (cpu)
• Jax version: 0.4.26
• JaxLib version: 0.4.26
• Using distributed or parallel set-up in script?: no

Who can help?

@sanchit-gandhi

Information

• [ ] The official example scripts
• [ ] My own modified scripts

Tasks

• [ ] An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• [ ] My own task or dataset (give details below)

Reproduction

Minimal reproduction:

import torch

model_outputs = [
    {
        'stride': [30, 0, 5],
        'tokens': torch.tensor([[
            50257, 50362, 8410, 7283, 0, 2329,
            8410, 7283, 0, 2094, 470, 1309,
            534, 10625, 307, 10625, 13, 34668,
            11, 345, 531, 9439, 11, 523,
            655, 8410, 7283, 0, 39134, 16592,
            10560, 3955, 50, 0, 7102, 5446,
            46, 0, 25848, 8410, 7283, 0,
            2773, 661, 4320, 1943, 981, 345,
            821, 8066, 7765, 510, 290, 670,
            1327, 379, 340, 13, 10528, 318,
            5340, 13, 50256
        ]]),
        'token_timestamps': torch.tensor([[
            0, 0, 0, 3.78, 4.22, 5.26, 6.04,
            6.54, 7, 7.94, 8.58, 8.58, 8.88, 9.16,
            9.54, 9.94, 10.6, 11.38, 11.88, 12.38, 12.44,
            12.62, 13, 13.36, 13.64, 14.24, 14.74, 15.12,
            15.4, 15.74, 16.1, 16.54, 16.54, 16.78, 17.08,
            17.2, 17.36, 17.56, 18.08, 18.58, 19.38, 19.88,
            22.54, 22.9, 23.24, 23.5, 24.14, 24.56, 24.7,
            24.94, 24.94, 25.18, 25.54, 25.72, 26.04, 26.34,
            26.46, 26.84, 27.04, 27.14, 27.54, 28.06, 29.92
        ]])
    },
    {
        'stride': [30, 5, 5],
        'tokens': torch.tensor([[
            50257, 50362, 2773, 661, 4320, 1943, 981,
            345, 821, 8066, 7765, 510, 290, 670,
            1327, 379, 340, 13, 10528, 318, 5340,
            13, 921, 815, 651, 284, 262, 966,
            810, 2687, 2073, 561, 11238, 290, 345,
            821, 407, 8066, 2245, 612, 13, 1400,
            11, 644, 389, 345, 4953, 329, 30,
            2141, 340, 0, 2329, 466, 340, 0,
            3363, 11, 345, 460, 0, 2329, 466,
            340, 0, 50256
        ]]),
        'token_timestamps': torch.tensor([[
            0, 0, 0, 2.92, 3.24, 3.5, 4.14,
            4.56, 4.7, 4.74, 4.92, 5.18, 5.54, 5.74,
            6.04, 6.34, 6.46, 6.84, 7.04, 7.18, 7.56,
            8.12, 9.68, 10.7, 10.88, 11.1, 11.24, 11.48,
            11.82, 12.46, 12.82, 13.2, 13.46, 13.72, 14.08,
            14.28, 14.34, 14.56, 14.82, 15.16, 15.72, 16.42,
            16.82, 16.86, 17, 17.1, 17.2, 17.56, 18.06,
            19.28, 19.6, 20.28, 21.96, 22.64, 24.28, 24.76,
            25.18, 25.56, 25.56, 25.84, 26.36, 27.12, 27.54,
            27.82, 28.16, 29.48
        ]])
    },
    {
        'stride': [23.7728125, 5, 0],
        'tokens': torch.tensor([[
            50257, 50362, 2329, 466,
            340, 0, 3363, 345,
            460, 0, 2329, 466,
            340, 0, 1002, 534,
            15867, 318, 3599, 625,
            11, 2245, 3501, 510,
            13, 50256
        ]]),
        'token_timestamps': torch.tensor([[
            0, 0, 0, 2.44, 4.3,
            5.04, 5.06, 5.56, 5.8, 6.32,
            7.12, 7.56, 7.8, 8.72, 10.04,
            12.96, 13.3, 13.44, 13.72, 13.98,
            14.86, 15.5, 16, 16.88, 17.76,
            20.9
        ]])
    }
]

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('onnx-community/whisper-tiny.en_timestamped')
tokenizer._decode_asr(model_outputs, return_timestamps='word', return_language=False, time_precision=0.02)

produces the following incorrect transcript:

(" DO IT! Just DO IT! Don't let your dreams be dreams. Yesterday, you said tomorrow, so just DO IT! MAKE YOUR DRIMS! CONTRO! JUST DO IT! Some people dream success while you're gonna wake up and work hard at it. Nothing is impossible. You should get to the point where anyone else would quit and you're not gonna stop there. No, what are you waiting for? Do it! Just do it! Yes, you can! Just do it! Yes you can! Just do it! If your tire is starting over, stop giving up.",
 {'chunks': [{'text': ' DO', 'timestamp': (0.0, 3.78)},
   {'text': ' IT!', 'timestamp': (3.78, 5.26)},
   {'text': ' Just', 'timestamp': (5.26, 6.04)},
   {'text': ' DO', 'timestamp': (6.04, 6.54)},
   {'text': ' IT!', 'timestamp': (6.54, 7.94)},
   {'text': " Don't", 'timestamp': (7.94, 8.58)},
   {'text': ' let', 'timestamp': (8.58, 8.88)},
   {'text': ' your', 'timestamp': (8.88, 9.16)},
   {'text': ' dreams', 'timestamp': (9.16, 9.54)},
   {'text': ' be', 'timestamp': (9.54, 9.94)},
   {'text': ' dreams.', 'timestamp': (9.94, 11.38)},
   {'text': ' Yesterday,', 'timestamp': (11.38, 12.38)},
   {'text': ' you', 'timestamp': (12.38, 12.44)},
   {'text': ' said', 'timestamp': (12.44, 12.62)},
   {'text': ' tomorrow,', 'timestamp': (12.62, 13.36)},
   {'text': ' so', 'timestamp': (13.36, 13.64)},
   {'text': ' just', 'timestamp': (13.64, 14.24)},
   {'text': ' DO', 'timestamp': (14.24, 14.74)},
   {'text': ' IT!', 'timestamp': (14.74, 15.4)},
   {'text': ' MAKE', 'timestamp': (15.4, 15.74)},
   {'text': ' YOUR', 'timestamp': (15.74, 16.1)},
   {'text': ' DRIMS!', 'timestamp': (16.1, 17.08)},
   {'text': ' CONTRO!', 'timestamp': (17.08, 18.08)},
   {'text': ' JUST', 'timestamp': (18.08, 18.58)},
   {'text': ' DO', 'timestamp': (18.58, 19.38)},
   {'text': ' IT!', 'timestamp': (19.38, 22.54)},
   {'text': ' Some', 'timestamp': (22.54, 22.9)},
   {'text': ' people', 'timestamp': (22.9, 23.24)},
   {'text': ' dream', 'timestamp': (23.24, 23.5)},
   {'text': ' success', 'timestamp': (23.5, 24.14)},
   {'text': ' while', 'timestamp': (24.14, 24.56)},
   {'text': " you're", 'timestamp': (24.56, 24.94)},
   {'text': ' gonna', 'timestamp': (24.94, 24.94)},
   {'text': ' wake', 'timestamp': (24.94, 25.18)},
   {'text': ' up', 'timestamp': (25.18, 25.54)},
   {'text': ' and', 'timestamp': (25.54, 25.74)},
   {'text': ' work', 'timestamp': (25.74, 26.04)},
   {'text': ' hard', 'timestamp': (26.04, 26.34)},
   {'text': ' at', 'timestamp': (26.34, 26.46)},
   {'text': ' it.', 'timestamp': (26.46, 27.04)},
   {'text': ' Nothing', 'timestamp': (27.04, 27.18)},
   {'text': ' is', 'timestamp': (27.18, 27.56)},
   {'text': ' impossible.', 'timestamp': (27.56, 29.68)},
   {'text': ' You', 'timestamp': (29.68, 30.7)},
   {'text': ' should', 'timestamp': (30.7, 30.88)},
   {'text': ' get', 'timestamp': (30.88, 31.1)},
   {'text': ' to', 'timestamp': (31.1, 31.24)},
   {'text': ' the', 'timestamp': (31.24, 31.48)},
   {'text': ' point', 'timestamp': (31.48, 31.82)},
   {'text': ' where', 'timestamp': (31.82, 32.46)},
   {'text': ' anyone', 'timestamp': (32.46, 32.82)},
   {'text': ' else', 'timestamp': (32.82, 33.2)},
   {'text': ' would', 'timestamp': (33.2, 33.46)},
   {'text': ' quit', 'timestamp': (33.46, 33.72)},
   {'text': ' and', 'timestamp': (33.72, 34.08)},
   {'text': " you're", 'timestamp': (34.08, 34.34)},
   {'text': ' not', 'timestamp': (34.34, 34.56)},
   {'text': ' gonna', 'timestamp': (34.56, 34.82)},
   {'text': ' stop', 'timestamp': (34.82, 35.16)},
   {'text': ' there.', 'timestamp': (35.16, 36.42)},
   {'text': ' No,', 'timestamp': (36.42, 36.86)},
   {'text': ' what', 'timestamp': (36.86, 37.0)},
   {'text': ' are', 'timestamp': (37.0, 37.1)},
   {'text': ' you', 'timestamp': (37.1, 37.2)},
   {'text': ' waiting', 'timestamp': (37.2, 37.56)},
   {'text': ' for?', 'timestamp': (37.56, 39.28)},
   {'text': ' Do', 'timestamp': (39.28, 39.6)},
   {'text': ' it!', 'timestamp': (39.6, 41.96)},
   {'text': ' Just', 'timestamp': (41.96, 42.64)},
   {'text': ' do', 'timestamp': (42.64, 44.28)},
   {'text': ' it!', 'timestamp': (44.28, 45.18)},
   {'text': ' Yes,', 'timestamp': (45.18, 45.56)},
   {'text': ' you', 'timestamp': (45.56, 45.84)},
   {'text': ' can!', 'timestamp': (45.84, 47.12)},
   {'text': ' Just', 'timestamp': (47.12, 47.54)},
   {'text': ' do', 'timestamp': (47.54, 47.82)},
   {'text': ' it!', 'timestamp': (44.3, 45.06)},
   {'text': ' Yes', 'timestamp': (45.06, 45.56)},
   {'text': ' you', 'timestamp': (45.56, 45.8)},
   {'text': ' can!', 'timestamp': (45.8, 47.12)},
   {'text': ' Just', 'timestamp': (47.12, 47.56)},
   {'text': ' do', 'timestamp': (47.56, 47.8)},
   {'text': ' it!', 'timestamp': (47.8, 50.04)},
   {'text': ' If', 'timestamp': (50.04, 52.96)},
   {'text': ' your', 'timestamp': (52.96, 53.3)},
   {'text': ' tire', 'timestamp': (53.3, 53.44)},
   {'text': ' is', 'timestamp': (53.44, 53.72)},
   {'text': ' starting', 'timestamp': (53.72, 53.98)},
   {'text': ' over,', 'timestamp': (53.98, 55.5)},
   {'text': ' stop', 'timestamp': (55.5, 56.0)},
   {'text': ' giving', 'timestamp': (56.0, 56.88)},
   {'text': ' up.', 'timestamp': (56.88, 60.9)}]})

(Notice at ~46 seconds, it goes back in time):

  {'text': ' Yes,', 'timestamp': (45.18, 45.56)},
   {'text': ' you', 'timestamp': (45.56, 45.84)},
   {'text': ' can!', 'timestamp': (45.84, 47.12)},
   {'text': ' Just', 'timestamp': (47.12, 47.54)},
   {'text': ' do', 'timestamp': (47.54, 47.82)},
   {'text': ' it!', 'timestamp': (44.3, 45.06)},
   {'text': ' Yes', 'timestamp': (45.06, 45.56)},
   {'text': ' you', 'timestamp': (45.56, 45.8)},
   {'text': ' can!', 'timestamp': (45.8, 47.12)},
   {'text': ' Just', 'timestamp': (47.12, 47.56)},
   {'text': ' do', 'timestamp': (47.56, 47.8)},
   {'text': ' it!', 'timestamp': (47.8, 50.04)},

For reference, this is the media I am transcribing.

Expected behavior

1. The transcript times should be increasing.
2. If you watch the video, it's clear that the repeated phrasing messes something up, duplicating this in the merged output.
3. Result should be something like:

   {'text': ' Do', 'timestamp': (39.28, 39.6)},
   {'text': ' it!', 'timestamp': (39.6, 41.96)},
   {'text': ' Just', 'timestamp': (41.96, 42.64)},
   {'text': ' do', 'timestamp': (42.64, 44.28)},
   {'text': ' it!', 'timestamp': (44.28, 45.18)},
   -  {'text': ' Yes,', 'timestamp': (45.18, 45.56)},
   -  {'text': ' you', 'timestamp': (45.56, 45.84)},
   -  {'text': ' can!', 'timestamp': (45.84, 47.12)},
   -  {'text': ' Just', 'timestamp': (47.12, 47.54)},
   -  {'text': ' do', 'timestamp': (47.54, 47.82)},
   -  {'text': ' it!', 'timestamp': (44.3, 45.06)},
   -  {'text': ' Yes', 'timestamp': (45.06, 45.56)},
   +  {'text': ' Yes', 'timestamp': (45.18, 45.56)},
   {'text': ' you', 'timestamp': (45.56, 45.8)},
   {'text': ' can!', 'timestamp': (45.8, 47.12)},
   {'text': ' Just', 'timestamp': (47.12, 47.56)},
   {'text': ' do', 'timestamp': (47.56, 47.8)},
   {'text': ' it!', 'timestamp': (47.8, 50.04)},

[xenova]

To help with debugging, here are the decoded outputs of each chunk:

for output in model_outputs:
  print(tokenizer.batch_decode(output['tokens']))

["<|startoftranscript|><|notimestamps|> DO IT! Just DO IT! Don't let your dreams be dreams. Yesterday, you said tomorrow, so just DO IT! MAKE YOUR DRIMS! CONTRO! JUST DO IT! Some people dream success while you're gonna wake up and work hard at it. Nothing is impossible.<|endoftext|>"]
["<|startoftranscript|><|notimestamps|> Some people dream success while you're gonna wake up and work hard at it. Nothing is impossible. You should get to the point where anyone else would quit and you're not gonna stop there. No, what are you waiting for? Do it! Just do it! Yes, you can! Just do it!<|endoftext|>"]
['<|startoftranscript|><|notimestamps|> Just do it! Yes you can! Just do it! If your tire is starting over, stop giving up.<|endoftext|>']

Indeed, the duplicated phrasing is at the word boundaries, so we can see where the algorithm messes up.