Incorrect Whisper long-form decoding timestamps

[Robinysh]

System Info

• transformers version: 4.42.4
• Platform: Linux-6.8.9-arch1-2-x86_64-with-glibc2.39
• Python version: 3.11.9
• Huggingface_hub version: 0.23.4
• Safetensors version: 0.4.3
• Accelerate version: not installed
• Accelerate config: not found
• PyTorch version (GPU?): 2.3.1 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?: No
• Using GPU in script?: Yes
• GPU type: NVIDIA GeForce RTX 3090

Who can help?

@Narsil @sanchit-gandhi

Information

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

Tasks

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

Reproduction

import numpy as np
import json
from transformers import pipeline, AutoModelForSpeechSeq2Seq, AutoProcessor
import torch
from datasets import load_dataset

device = "cuda"
torch_dtype = torch.bfloat16
# model_id = "openai/whisper-large-v3"
model_id = "distil-whisper/distil-large-v3"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=False, use_safetensors=True
)
model.to(device)
processor = AutoProcessor.from_pretrained(model_id)
pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=None,
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset(
    "hf-internal-testing/librispeech_asr_dummy", "clean", split="validation"
)
sample = dataset[0]["audio"]
sample = np.concatenate([sample["array"]] * 10)

results = pipe(
    sample,
    return_timestamps=True,
    generate_kwargs={
        "language": "english",
    },
)
print(json.dumps(results, indent=4))

Output

{
    "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel. Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel. Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome His Gospel. Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel. Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel. Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel. Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel. Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel. Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel. Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel.",
    "chunks": [
        {
            "timestamp": [
                0.0,
                6.5
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel."
        },
        {
            "timestamp": [
                6.5,
                12.5
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel."
        },
        {
            "timestamp": [
                12.5,
                18.24
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome His Gospel."
        },
        {
            "timestamp": [
                18.24,
                24.0
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel."
        },
        {
            "timestamp": [
                24.0,
                29.84
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel."
        },
        {
            "timestamp": [
                0.0,
                4.7
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel."
        },
        {
            "timestamp": [
                5.76,
                10.54
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel."
        },
        {
            "timestamp": [
                11.6,
                16.4
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel."
        },
        {
            "timestamp": [
                17.46,
                22.26
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel."
        },
        {
            "timestamp": [
                23.12,
                28.12
            ],
            "text": " Mr. Quilter is the Apostle of the Middle Classes, and we are glad to welcome his gospel."
        }
    ]
}

Expected behavior

Currently the timestamp resets to zero after 30s of audio. I expect the timestamps to increase monotonically.

[Robinysh]

Additionally, the bug does not happen when I add chunk_length_s=30 to pipe(), i.e.

results = pipe(
    sample,
    chunk_length_s=30,
    return_timestamps=True,
    generate_kwargs={
        "language": "english",
    },
)

However this workaround is not applicable to me because I also would like to supply the compression_ratio_threshold argument to generate_kwargs, and that is not supported with short-form transcription.

[amyeroberts]

cc @kamilakesbi

[kamilakesbi]

Thanks for opening this issue @Robinysh!

This is indeed a problem, I'll open an issue to solve it!

Note that we're working on unifying short form and long form generation with PR #30984. Once merged you should be able to use compression_ratio_threshold with short-form transcription :)

[kamilakesbi]

Hi @Robinysh,

You could use this workaround before we properly integrate the solution in Transformers:

import numpy as np
import json
from transformers import AutoProcessor, WhisperForConditionalGeneration
import torch
from datasets import load_dataset

device = "cuda"
torch_dtype = torch.bfloat16

processor = AutoProcessor.from_pretrained("distil-whisper/distil-large-v3")
model = WhisperForConditionalGeneration.from_pretrained("distil-whisper/distil-large-v3", torch_dtype=torch.float16)
model = model.to("cuda")

dataset = load_dataset(
    "hf-internal-testing/librispeech_asr_dummy", "clean", split="validation"
)
sample = dataset[0]["audio"]
sample = np.concatenate([sample["array"]] * 10)

inputs = processor(sample, return_tensors="pt", truncation=False, sampling_rate=16_000)
inputs = inputs.to("cuda", torch.float16)

output = model.generate(**inputs, return_timestamps=True, return_segments = True)

result = processor.batch_decode(output['sequences'], skip_special_tokens=True, output_offsets = True)

for i in range(len(result[0]['offsets'])):
   result[0]['offsets'][i]['timestamp'] = (output['segments'][0][i]['start'].item(), output['segments'][0][i]['end'].item())

print(json.dumps(result, indent=4))

Explanation:

When performing long form generation with Whisper, the right utterance level timestamps are returned as output to generate when we specify return_segments = True and return_timestamps=True.

The problem arise at the decoding level: batch_decode currently doesn't support the output to long form generation (segments). When specifying output_offsets, it indeed outputs the wrong timestamps you previously obtained.

One simple solution is to replace the obtained timestamps with the ones stored in output[segments], which I do with these lines:

for i in range(len(result[0]['offsets'])):
   result[0]['offsets'][i]['timestamp'] = (output['segments'][0][i]['start'].item(), output['segments'][0][i]['end'].item())

cc @sanchit-gandhi @ylacombe ( We should integrate this properly in batch_decode and also handle it in the automatic speech recognition pipeline, I'll open a PR for that :) )