ASR pipeline for Estonian speech recognition

This repository contains two separate Nextflow pipelines:

• transcribe.nf is the default CPU-only Kaldi pipeline.
• transcribe-whisper-gpu.nf is a Whisper/CUDA pipeline and requires an NVIDIA GPU.

The pipelines intentionally use separate entrypoints, configs, and Dockerfiles because their runtime dependencies are different.

Installation

Install Java and Nextflow. Nextflow 23.0.0 or later is required.

wget -qO- https://get.nextflow.io | bash
chmod +x nextflow
mkdir -p $HOME/.local/bin/
mv nextflow $HOME/.local/bin/

Docker is recommended for both pipelines.

CPU/Kaldi Pipeline

Pull the CPU image:

docker pull alumae/est-asr-pipeline:0.1.4

Run the default CPU pipeline:

nextflow run transcribe.nf --in /path/to/audio.mp3

On machines with limited RAM, run serially with one CPU and a Docker memory cap:

nextflow run transcribe.nf -profile lowmem --in /path/to/audio.mp3

CPU and memory configuration

The CPU/Kaldi workflow uses two shared parameters for the heavier steps:

• --nthreads <N>: number of CPU threads requested by speaker ID and decoding.
• --process_memory <SIZE>: memory requested by diarization, speaker ID, decoding, and RNNLM rescoring.

For example, to request four CPU threads and 8 GB for the main processing steps:

nextflow run transcribe.nf --in /path/to/audio.mp3 --nthreads 4 --process_memory 8GB

Some lightweight steps have fixed resource settings in transcribe.nf, for example to_wav uses 500MB and 2 CPUs, language_id uses 4GB, mfcc uses 1GB, lattice2ctm uses 4GB and 2 CPUs, and punctuation uses 4GB. The --process_memory parameter does not override those fixed per-process settings.

The bundled lowmem profile is useful on small machines:

nextflow run transcribe.nf -profile lowmem --in /path/to/audio.mp3

It sets the workflow to run one task at a time, uses one thread for threaded Kaldi steps, lowers process_memory to 3GB, and enables Docker with a hard container limit:

process.maxForks = 1
executor.queueSize = 1
params.nthreads = 1
params.process_memory = "3GB"
docker.runOptions = "--memory=5g --memory-swap=5g --cpus=1"

Nextflow cpus and memory directives are resource requests used for scheduling. With the local executor, they are not always hard operating-system limits. When running through Docker, docker.runOptions is the part that enforces actual container-level CPU and memory caps.

For custom limits, create a small config file, for example resources.config:

params {
    nthreads = 4
    process_memory = '8GB'
}

process {
    maxForks = 1

    withName: 'one_pass_decoding' {
        cpus = 4
        memory = '8GB'
    }

    withName: 'speaker_id' {
        cpus = 4
        memory = '8GB'
    }

    withName: 'punctuation' {
        memory = '6GB'
    }
}

docker.runOptions = '--memory=10g --memory-swap=10g --cpus=4'

Run it with:

nextflow run transcribe.nf -c resources.config --in /path/to/audio.mp3

By default, results are written to results/<audio-basename>/:

result.ctm  result.json  result.srt  result.trs  result.with-compounds.ctm  result.txt

The CPU pipeline writes plain transcript text to result.txt.

Whisper/CUDA Pipeline

The Whisper pipeline requires an NVIDIA GPU, CUDA-compatible drivers, and the NVIDIA container runtime when using Docker.

Pull the GPU image:

docker pull europe-north1-docker.pkg.dev/speech2text-218910/repo/est-asr-pipeline:1.1b

Run the Whisper pipeline:

nextflow -C nextflow.whisper-gpu.config run transcribe-whisper-gpu.nf -profile docker --in /path/to/audio.mp3

For SGE:

nextflow -C nextflow.whisper-gpu.config run transcribe-whisper-gpu.nf -profile docker,sge --in /path/to/audio.mp3

For SLURM:

nextflow -C nextflow.whisper-gpu.config run transcribe-whisper-gpu.nf -profile docker,slurm --in /path/to/audio.mp3

By default, Whisper results are written to results/.

The Whisper pipeline writes speaker-prefixed transcript lines to result.txt.

Parameters

Common parameters:

• --in <filename>: audio or video file to transcribe.
• --out_dir <path>: output directory.
• --do_speaker_id true|false: include speaker identification.
• --do_language_id true|false: filter non-Estonian speech segments.

CPU-only parameter:

• --do_punctuation true|false: run punctuation restoration.

Whisper-only parameter:

• --in_file_list <filename>: file containing input paths, one per line.

Configuration

CPU/Kaldi configuration lives in nextflow.config.

Whisper/CUDA configuration lives in nextflow.whisper-gpu.config.

The Dockerfiles are separate:

• Dockerfile: CPU/Kaldi image recipe.
• Dockerfile.whisper-gpu: Whisper/CUDA image recipe.

Nextflow Options

Useful standard Nextflow options:

• -with-report: write an HTML execution report.
• -with-trace: write a machine-readable execution trace.
• -with-dag <filename.png>: write a workflow graph.