Solo speech ASR reference: inspect the waveform, automatic text and cue boundaries

The exact public reference and evidence state

Inspect one real 13.525333-second public MP3 excerpt: mono speech at 48 kHz, with an automatic transcript and a cue file whose roles remain distinct.

These files were already published on Fiund’s specialist websites as owner-authorized editorial media. This study documents the exact public reference and its limits. It does not advertise an accepted commercial training collection, grant downstream use rights or establish permissions for every person or contribution depicted.

Use the measured reference manifest to inspect the delivered filenames, measured properties and SHA-256 values. The manifest identifies this small public reference; it does not describe the size of an uninspected private archive.

What the published speech file measures

Direct inspection reports an MP3 audio stream with one channel and a 48,000 Hz sample rate. The container duration is 13.525333 seconds, and the published file contains 162,360 bytes. Those are properties of this short public derivative, not of a complete podcast source or a corpus.

A single mono export does not preserve evidence of original microphone separation. It also cannot establish a speaker distribution, recording-condition diversity or a domain-wide speech-recognition evaluation set.

Automatic text is not a gold reference

The transcript was produced automatically and lightly normalized for punctuation and capitalization. Prior source preparation recorded a matching second automatic transcription of the selected excerpt. That consistency is useful preparation evidence, but it is not independent human adjudication.

A reliable recognition evaluation needs an explicit transcript policy and a review process. Names, contractions, hesitations, partial words and punctuation can be treated differently by different tasks. This short example does not provide a validated word error rate or prove the performance of any model.

Keep cue timing and recognition evaluation separate

The VTT cue file provides display timing for the public text. A cue interval is not a word-level forced alignment, a phoneme boundary or a verified voice-activity label. The encoding duration can also include codec padding relative to a nominal source trim.

For an ASR pilot, define the audio version, scoring text normalization and error-counting policy together. Retain the original text and any normalized scoring form so a change of convention does not appear to be a model improvement.

• Inspect the exact decoded audio and its duration before scoring.
• Use qualified reference review for ambiguous words.
• Do not infer diarization quality from a solo mono excerpt.
• Treat AIPodcast and Podcast Datasets versions as one underlying source.

From the reference to a real project

A useful project request starts with the actual task and the evidence it needs. Ask which source units, permissions, annotations and review steps would be required for your application rather than assuming that a public example establishes inventory.

The public manifest can be used to reconcile file identity. A matching checksum means the bytes match that reference; it does not establish source ownership, consent, completeness, annotation accuracy or fitness for a model task. Those questions need their own evidence.

Contact Fiund with the desired source characteristics and the decision your model must make. Confirm sample access, source scope, preparation and commercial terms for the selected project before making a larger collection or delivery commitment.

Questions buyers ask