PVSpeech

Clean read Nepali speech for evaluating speech recognition, from Ampixa Labs: 50 clips, 4.2 minutes, 16-bit PCM mono at 24 kHz, from two adult speakers who consented to public release.

Evaluation only. Please do not train on this data.

Contents

• audio/: one WAV per clip.
• references.json: each clip's reference text, speaker, duration, sha256, and whether it is in the 93-sentence panel (in_pv93).

What this set is

The full PVSpeech set is 100 sentences from four speakers. This release holds the 50 sentences from speakers pv01 and pv04. The other two speakers' recordings are not released and remain a private holdout, so scores on this release are not the same as scores on the full set.

Speakers read Wikipedia and news sentences in a neutral style. Speaker IDs are anonymous.

PVSpeech helped select Rani's checkpoint and tune its language model, so Rani's result on it is not an untouched test.

Results and scorer: https://github.com/Ampixa/rani-asr. Contributors may withdraw their recordings at any time by contacting Ampixa Labs.