taskworker-transcribe

Worker for taskbridge which can handle tasks of type transcribe.

Result format

When calling the TaskBridge /api/tasks/complete/:id API, the following JSON structure is sent to the endpoint.

{
  "result" : {
    "language" : "en",
    "texts" : [
      {
        "start" : 0.0,
        "end" : 1.0,
        "text" : "Line 1"
      },
      {
        "start" : 1.0,
        "end" : 2.0,
        "text" : "Line 2"
      }
    ],
    "device" : "cuda",
    "duration" : 12,
    "repository" : "https://github.com/hilderonny/taskworker-transcribe",
    "version" : "1.1.0",
    "library" : "fasterwhisper-0.8.15",
    "model" : "large-v2"
  }
}

Property	Description
language	Detected langugage in the file depending on the first seconds as 2 digit ISO code
texts	Array of detected text chunks. Normally the text is splitted by sentences
texts.start	Start second of the chunk
texts.end	End second of the chunk
texts.text	Text content of the chunk
device	cuda for GPU processing and cpu for CPU processing
duration	Time in seconds for the processing
repository	Source code repository of the worker
version	Version of the worker
library	Library used to perform transcription
model	AI model used for transcription

Installation on Windows

First install Python 3.11. Currently I cannot get the actual faster-whisper 1.0.3 implementation to work with CUDA 12. So I need to use version 0.9.0 which is installable only with Python 3.11. Then run the following commands in the folder of the downloaded repository.

python3.11 -m venv python-venv
python-venv\Scripts\activate
pip install faster-whisper==0.9.0

Next you need to copy cudnn_ops_infer64_8.dll, cudnn_cnn_infer64_8.dll, cublas64_11.dll, cublasLt64_11.dll and zlibwapi.dll (Can be found on NVIDIA websites) into the directory python-venv/Lib/site-packages/ctranslate2.

Installation as service under Linux

sudo apt install -y git python3.11-env ocl-icd-libopencl1 nvidia-cuda-toolkit nvidia-utils-510-server nvidia-utils-535-server
python3.11 -m venv python-venv
source python-venv/bin/activate
pip install faster-whisper==0.9.0 nvidia_cublas_cu11==11.11.3.6 nvidia_cudnn_cu11==9.4.0.58

Adopt the shell script translate.shto your needs and create SystemD config files (if you want tu run the worker as Linux service).

/etc/systemd/system/taskworker-transcribe.service:

[Unit]
Description=Task Worker - Audio Transcriber

[Service]
ExecStart=/taskworker-transcribe/transcribe.sh
Restart=always
User=user
WorkingDirectory=/taskworker-transcribe/

[Install]
WantedBy=multi-user.target

Finally register and start the services.

chmod +x ./transcribe.sh
sudo systemctl daemon-reload
sudo systemctl enable taskworker-transcribe.service
sudo systemctl start taskworker-transcribe.service

Running

Running the program the first time, ai models with about 4 GB (depending on the used model) get downloaded automatically.

python transcribe.py --taskbridgeurl http://192.168.178.39:42000/ --device cuda --worker ROG --model large-v2

Literature

1. https://github.com/SYSTRAN/faster-whisper