Batch Transcription

Transcribe speech 100x faster and 100x cheaper with Modal and NeMo ASR models.

Setup

• Clone this repo
• Install uv
• Build the virtual environment: uv sync
• Setup your Modal account (modal setup)
• Add a Modal API token to your environment if necessary (modal token new)

Models

Any NeMo ASR model should work. Though it may be necessary to handle model specific kwargs to transcribe.

We've tested the following models:

• nvidia/parakeet-tdt-0.6b-v2 (default): Hyperfast, English only transcription
• nvidia/canary-1b-flash: Just regular fast multilingual transcription

Model weight caching

The first run for each model will incur a small latency cost to download the weights to cache. Subsequent runs will load the weights from the Modal Volume: transcription-models.

Usage

Download ESB Test Datasets

First stage the data (one-time setup) on the Modal Volume: transcription-datasets:

modal run -m run::stage_data

This downloads audio files from the HuggingFace ESB test subsets: AMI, Earnings22, GigaSpeech, LibriSpeech (clean/other), SPGISpeech, TEDLIUM, VoxPopuli.

Run Batch Transcription

modal run -m run::batch_transcription

Or run with arguments:

modal run -m run::batch_transcription \
  --model_id nvidia/parakeet-tdt-0.6b-v2 \
  --gpu-type L40S \
  --gpu-batch-size 128 \
  --num-requests 25 \
  --job-id my-transcription-job

Configuration Options

Argument	Default	Description
--model_id	nvidia/parakeet-tdt-0.6b-v2	NeMo ASR model identifier
--gpu-type	L40S	GPU type for transcription function
--gpu-batch-size	128	Number of audio files per GPU batch
--num-requests	25	Number of parallel Modal function calls
--output-path	results	Path for results directory
--job-id	Auto-generated if not provided	Job identifier

Output

Results are saved to the Modal Volume, transcription-results, in two formats:

1. Summary: /results_summaries/results_summary_{job_id}.csv

   • Aggregated metrics (WER, RTFX, timing)

2. Detailed: /results/{job_id}.csv

   • Individual transcriptions, ground truth, dataset info

Metrics

• WER: Word Error Rate (%) calculated using normalized text for one request
• RTFX: Real-time factor (audio duration / processing time) for one request
• Total Runtime: End-to-end job execution time for whole job

normalizer

The normalizer module in this repo used to process text and score WER is pulled from the HuggingFace ASR Leaderboard.