Timething is a library for aligning text transcripts with audio. You provide an audio file, as well as a text file with the complete text transcript. Timething will output a list of time-codes for each word and character that indicate when this word or letter was spoken in the audio you provided. Timething strives to be fast and accurate, and can run on both GPUs or CPUs.

Timething uses powerful Wav2Vec based speech recognition models hosted by the Hugging Face AI community. The approach is described in this PyTorch Tutorial, as well as in this paper.

To install Timething, you'll need an installation of Python 3.7 or 3.8. You can then install it using pip:

pip install timething

There are many cases where you might want to align long audio content with a corresponding transcript. For example you might want to align a podcast with its transcription. The episode might be a few hours long, and the transcription is given by the podcaster. In another case you might want to align an audio book with it's written text.

timething align-long \
    --audio-file fixtures/audio/keanu.mp3 \
    --transcript-file fixtures/keanu.cleaned.txt \
    --alignments-dir aligned \
    --batch-size 10 \
    --n-workers 5

Timething can align a dataset of utterance level audio snippets with their text transcriptions. This is particularly useful in a machine learning setting, where Timething can be used to clean up datasets.

Timething currently expects to find a folder containing one or more chapters in the following form:

└── dir/
    ├── text.csv
    ├── aligned/
    └── audio/
        ├── chapter01.mp3
        ├── chapter02.mp3
        └── chapter03.mp3

Timething can process many audio formats, including MP3, WAV, FLACC and OGG/VORBIS.

The file text.csv should contain one entry per wav file in the following format:

audio/chapter01.mp3|The transcript for chapter01 on a single line here
audio/chapter02.mp3|The transcript for chapter02 on a single line here
audio/chapter03.mp3|The transcript for chapter03 on a single line here

You can now run Timething on your CPU or GPU, for example:

timething align-short --metadata text.csv --alignments-dir aligned

You can also specify more options, e.g.:

timething align-short \
  --language german \
  --metadata text.csv \
  --alignments-dir aligned \
  --batch-size 8 \
  --n-workers 8

Run timething --help for a full description.

Results will be written into the given folder, e.g. aligned. They will be written into a single json file named after each audio id. Each file will contain the character level and the word level alignments. For word level alignments, each word will have the starting time in seconds, the ending time in seconds, the confidence level for that word and the word label. Character level alignments have the corresponding results.

You can find an example dataset with alignments output in fixtures/. Here's what the alignment for "one.mp3", which contains only the word "one", looks like:

{
    "n_model_frames": 72,
    "n_audio_samples": 23392,
    "sampling_rate": 16000,
    "chars": [
        {
            "label": "O",
            "start": 0.5888611111111111,
            "end": 0.6497777777777777,
            "score": 0.9999777873357137
        },
        {
            "label": "n",
            "start": 0.6497777777777777,
            "end": 0.7106944444444444,
            "score": 0.99994424978892
        },
        {
            "label": "e!",
            "start": 0.7106944444444444,
            "end": 0.731,
            "score": 0.9999799728393555
        }
    ],
    "chars_cleaned": [
        {
            "label": "o",
            "start": 0.5888611111111111,
            "end": 0.6497777777777777,
            "score": 0.9999777873357137
        },
        {
            "label": "n",
            "start": 0.6497777777777777,
            "end": 0.7106944444444444,
            "score": 0.99994424978892
        },
        {
            "label": "e",
            "start": 0.7106944444444444,
            "end": 0.731,
            "score": 0.9999799728393555
        }
    ],
    "words": [
        {
            "label": "One!",
            "start": 0.5888611111111111,
            "end": 0.731,
            "score": 0.9999637263161796
        }
    ],
    "words_cleaned": [
        {
            "label": "one",
            "start": 0.5888611111111111,
            "end": 0.731,
            "score": 0.9999637263161796
        }
    ]
}

Once you've run alignment, you can cut your files down to smaller files and write the results into a new folder. For example, if you don't want any of your recordings to exceed 8 seconds, then you can create a new directory and re-cut your data into it like this:

timething recut \
  --from-metadata text.csv \
  --to-metadata ~/smaller-recordings/text.csv \
  --alignments-dir alignments \
  --cut-threshold-seconds 8.0

Results in this example are written into ~/smaller-recordings.

Currently supported languages can be found in models.yaml. This currently includes English, German, Dutch, Polish, Italian, Portuguese, Spanish, French, Russian, Japanese, Greek and Arabic models. We have only tested the German model so far.

Due to the large number of CTC speech models available on the Hugging Face AI community, new languages can be easily added to Timething. Alternatively, Wav2Vec can be fine-tuned as described here, using any of the Common Voice languages, 87 at the time of writing.

Support for text cleaning is currently minimal, and may need to be extended for new languages.

There are many mature libraries that can already do forced alignment like Timething, e.g. the Montreal forced aligner or Aeneas. One list of tools is maintained here.

Thanks to why do birds for allowing the initial work on this library to be open sourced.