Public audio archives
Make permitted public recordings easier to search. Review the recording rights, speaker consent and transcript accuracy before expanding the dataset.
Research implementation · GPU testing pending
GPU audio transcription with measurable quality.
Turn an authorised speech recording into text for a searchable archive, subtitles or a transcript review workflow. The transcription kit evaluates short English clips offline; it measures errors against known transcripts rather than treating a plausible sentence as proof of accuracy.
Create an account and describe your workload →What this batch could help with
Transcript preparation
Create a first draft of a transcript that a person checks. Names, numbers, accents and technical terms can need correction even when the output reads fluently.
Speech model evaluation
Use a small known-text dataset to compare CPU and GPU execution, inference time and word error rate. Keep the same model files and decoding configuration in both runs.
How the research workload runs
- Prepare reviewed fixtures
Use public or synthetic English speech with an expected transcript. Record a SHA256 hash for each 16 kHz mono 16-bit PCM WAV file.
- Use a staged, pinned model
Run Whisper tiny.en from a local directory with a reviewed immutable revision and file-hash manifest. The agent permits safetensors weights and built-in model code; it makes no downloads.
- Review the transcript
Measure word error rate against the known text. Generate a separate CPU reference, compare the transcripts, then review errors and hallucinations before accepting any larger use case.
Inputs and limits
The first kit supports 1–16 English clips, each 1–30 seconds, with at most four minutes of audio per invocation. It uses FP32 greedy decoding with a 128-token output cap. Long recordings, diarisation, translation and live captions are outside this first implementation.
Software requirements
An existing CUDA environment with PyTorch, Transformers and NumPy, plus a separately reviewed local Whisper tiny.en model and original/public fixtures. The kit neither installs dependencies nor changes drivers. Missing files or mismatched hashes stop the run.
Results and checks
Each JSON result contains the fixture hash, original duration, expected text, generated transcript, inference time and word error count. The aggregate report includes word error rate, model revision, all reviewed model-file hashes, load time and peak PyTorch allocation.
Word error rate counts substitutions, deletions and insertions relative to a known transcript. CPU/GPU agreement checks repeatability, but two runs can agree on an incorrect transcript. Human review remains part of the acceptance decision; no general accuracy percentage is promised.
When a GPU is useful
A GPU can help with inference, but short clips may spend much of their time loading the model and preparing features. Report model load separately and compare like-for-like. A tiny model offers a bounded starting point; it does not guarantee acceptable accuracy for your domain.
Useful future quote units are accepted audio minutes with a defined language, recording format and quality target. Include preprocessing, model load, transcript review, retries and delivery in the comparison.
No published performance or price is established for this new workload. The embeddings benchmark does not predict its speed or accuracy.
Try the controlled research kit
The implementation runs locally in an existing reviewed environment. It uses no IdleGPU token, creates no customer jobs, installs nothing and awards no badge. GPU compatibility and useful output quality still need a real-device test. Customer dispatch is not enabled for this workload.
Before a marketplace launch, this workload needs independent correctness checks, interruption/retry testing, a reviewed data path, workload-specific verification and an agreed acceptance rule.
Common questions
Which speech languages are supported by the first kit?
The initial runner uses Whisper tiny.en for short English recordings. It does not claim multilingual transcription or translation support.
Can I use private meeting recordings?
Private recordings and personal information are outside the current IdleGPU public-data pilot. Use original synthetic speech or authorised public recordings with reviewed permissions.
Does a low word error rate guarantee a correct transcript?
No. A score on a small test set does not establish accuracy for every accent, topic or recording. Review names, numbers, hallucinations and domain-specific mistakes separately.
Explore related GPU workloads
Text embeddings → GPU image processing → GPU rendering →