Datasetpublic

sfsmcnulty/stt_test_audio

Mobile Voice Platform STT Test Audio Versioned benchmark audio for the MobileVoicePlatform-Android SampleApp. The repository intentionally contains two views of the same clips: data/ is an AudioFolder-compatible view with metadata.csv for Hugging Face tooling. packs/ contains checksum-pinned ZIPs optimized for bounded download and validation on Android. catalog.json is the machine-readable index used to discover datasets, categories, languages, checksums, clip references, and… See the full description on the dataset page: https://huggingface.co/datasets/sfsmcnulty/stt_test_audio.

sourceHugging Faceupdated 14d agoView on Hugging Face

0likes93downloads

Dataset Card

Mobile Voice Platform STT Test Audio

Versioned benchmark audio for the MobileVoicePlatform-Android SampleApp. The repository intentionally contains two views of the same clips:

—data/ is an AudioFolder-compatible view with metadata.csv for Hugging Face tooling.
—packs/ contains checksum-pinned ZIPs optimized for bounded download and validation on Android.
—catalog.json is the machine-readable index used to discover datasets, categories, languages, checksums, clip references, and mobile pack paths.
—asr-results-v1.json contains versioned transcripts and scores for Parakeet TDT v3 and Nemotron 3.5, plus Whisper baselines where available.

Dataset ID	Display name	Category	Audio type	Clips	Seconds
`general-multilingual`	General multilingual	`general`	`synthetic`	56	3670.5
`medical-synthetic`	Medical synthetic	`medical`	`synthetic`	36	530.0
`clinical-clones`	Clinical voice clones	`medical`	`synthetic`	500	7458.1

Mobile consumption

Fetch catalog.json, select a pack by dataset_id and language, then download its path from:

https://huggingface.co/datasets/sfsmcnulty/stt_test_audio/resolve/main/<path>

Verify byte_size and sha256 before opening a pack. Pack contents retain the voice_studio_clip_pack_v1 metadata consumed by the SampleApp.

clinical-clones is available through the AudioFolder view only. Its 500 clips exceed the mobile pack size limit, so it is intentionally absent from catalog.json and packs/.

Adding a dataset

Add one declarative entry to tools/sttbench/huggingface_dataset.json, supply its source archive to build_huggingface_dataset.py, regenerate, run the tests, and upload the generated directory. Dataset IDs are stable API identifiers; categories (for example medical) and audio types (for example human or synthetic) are independent filtering dimensions.

Provenance and use

The general corpus, synthetic medical corpus, and VoxCPM clinical voice clones were generated for internal STT evaluation. Human medical-domain recordings are intentionally excluded. Source archives do not include a formal license; verify distribution rights before mirroring or commercial reuse. Stored Whisper scores are directional references because scorer provenance is not included.