iamthen0ise/soundsplit

Sound Split and ASR

Split soundfiles by silence / low noise. Processing ASR via Yandex Cloud or Google Cloud. Compare ASR Results with text.

Requirements:

python 3.7
ffmpeg

Installation:

virtualenv venv -p python3 \
&& source venv/bin/activate \
&& pip install -r requirements.txt

RUN:

Audio Split

usage: split.py [-h] [-i INPUT_FILE] [-o OUTPUT_DIR] [-m METHOD] [-p PREFIX]
                [-fl FRAME_LENGTH] [-fs FRAME_SHIFT] [-l LIMIT]
                [-sr SAMPLERATE] [-q Q_FACTOR]

            Split audio files by chosen <method>.

            If method is `ina`, then using Ina Speech Segmenter.
            If method is `rms` or not specified, using RMS energy and Zero-Crossing.

            Ina provides more accurate results thought could be slow and cannot be configured.
            RMS method could be configured via params, but may return some broken audio chunks.

            Frame length is length of audio sample window
            Frame shift is a distance between audio samples

            Q-Factor is a factor used for smoothing RMS and Zero-Crossing peak values.
            e.g. if RMS=0.5 and Q-Factor=0.8 the resulting RMS would be 0.5*0.8

            Limit is a length of audio that should be splitter from start.

Example:

python src/split.py --input-file <input file path> \
--output-dir <output dir path> \
--prefix <prefix> \
--method rms
--frame-length 500 \
--frame-shift 50 \
--limit 300 \
--samplerate 44100 \
--q-factor 0.7

ASR

usage: asr.py [-h] [-i INPUT_DIR] [-ll LANGUAGE] [-l LIMIT] [-j JSONFILE]
              [--iam IAM] [--folder-id FOLDER_ID]

            Process ASR for audio files.

            ** CREDENTIALS **

            Please specify `--iam` and `--folder-id` for Yandex Services.
            See https://cloud.yandex.ru/docs/iam/operations/iam-token/create for reference.

            For Google Speech To Text use environmental variables and config as described here —
            https://cloud.google.com/speech-to-text/docs/libraries#linux-or-macos

Example:

python src/asr.py \
--input-dir <dir with splitted chunks> \
--iam <Yandex cloud iam token> \
--folder-id <Yandex cloud folder id> \
--language ru-RU \
--limit 50 \
--jsonfile <path to resulting json file>

TEXT EVALUATION

usage: text_eval.py [-h] [-i TEXT_INPUT] [-j JSONFILE] [-q Q_FACTOR]
                    [-qmax Q_FACTOR_MAX] [-qstep Q_FACTOR_STEP]

        Find similar text.

        Q-Factor is Levenshtein Distance value.
        Default Q-Factor used for retrieve text chunks from source text file.
        If chunks cannot be retrieved, Q-Factor increases step by step by value of Q-Factor Step
        Until reach Maximal Q-Factor.

        Using default values of Q-Factor higher than 7 may slowdown the script.

Example:

python src/text_eval.py \
--text-input <file with text source path> \
--jsonfile <result json file path \
--q-factor 3 \
--q-factor-max 70 \
--q-factor-step 5