Meeting: Clarification and concerns regarding Garchen Rinpoche's STT Model PRD

🧩 SIG Garchen Rinpoche Speech SIG
1

Attendees: [@DavidYesheNyima - Champion, @Ganga_Gyatso - developer, @Lhakpa_Wangyal - coordinator ]

Date: July 24, 2025

Agenda:

  • Clarify questions about the updated PRD prepared by Mr. David Newman.
  • Understand the expected final output for Garchen Rinpoche’s speech transcription.
  • Discuss whether the 5% Character Error Rate (CER) goal is realistic.

Key Discussion Points

  • Mr. Ganga Gyatso outlined about the completion of tests done with base model, and HEGR’s STT model trained with 5 hours.

  • If training continues in a straight line, he mentioned that we might achieve 5% CER with 27 hours of data. He emphasised that in reality progress might be non-linear and future improvements may slow down.

  • With reference to next possible steps, Mr. Ganga Gyasto added that if CER stops improving with more data, he suggested for reviewing CER at 10, 15, 20, and 25 training hours.

  • There was discussion about data Quality which affects results and suggested selection of high quality data.

  • It was also discussed that editing speed without model was observed 500 Tibetan syllables/hour approximately. In contrast, it was found ~1000 syllables/hour (2× faster) with fine tuned model.

  • Mr. David Newman discussed that the final model should be able to transcribe only speech of His Eminence Garchen Rinpoche to automatically produce timestamped transcripts in WebVTT format.

  • Mr. Ganga Gyatso mentioned that the HEGR’s STT model will work via API on Hugging Face with WebVTT output. For diarization, he specified that he has to train a model that will work together with the STT Model linked with the API on Hugging face.

  • It was discussed that target of 5% CER is flexible which may be updated based on real results.

  • Mr. Ganga Gyatso discussed about the progress report of the project using GitHub

  • Discussed about the need to update all the workflow in the Github.

  • Outlined about completing Project requirement document.

Action Items

  • @Ganga_Gyatso [Train up to 10 hours and check progress.] - Due: [by the end of this week]
  • @username**:** [Another task] - Due: [YYYY-MM-DD]

Decisions Made

  1. Decision: To Continue weekly progress updates and meeting whenever there is possibility of time.
  2. Decision: 5% CER is flexible which may be updated based on real results.
  3. Decision: Updating of the PRD by Mr. David Newman.