Meeting: Work progress update

🧩 SIG Garchen Rinpoche Speech SIG
1

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

Date: August 05, 2025

Agenda:

  • Clarify questions from by Mr. David Newman;
    1. Site for viewing the edited transcripts for the corresponding audio files
    2. Optimum model for continuing with fine tuning as stated in the conclusion of an article by Ganga Gyatso.
    3. To clarify concerns from Mr. David Newman regarding transcript cleanup and translation.
  • General aspects of the project like work progress, fine tuning and any additional

Key Discussion Points

  • Mr. David shared a concern about how the 10 hours of audio were fine-tuned, based on the conclusion of an article by Ganga Gyatso on the OpenPecha forum.
  • Mr. Ganga Gyatso responded by stating the comparative results of the fine tuning using base model of Openpecha and the already (five hours of audio data) fine tuned Garchen Rinpoche’s STT model. He stated the need of more training data to get better result by explaining with the comparative data from his article [ Garchen Rinpoche CER Analysis and Comparison].
  • Additionally, Ganga Gyatso la explained a smart and cost-effective Garchen Rinpoche’s audio extraction method using voice print technology to identify. He added that His Eminence Garchen Rinpoche’s voice is extracted from multi-speaker audio files by the technology.
  • Mr. Ganga Gyatso demonstrated transcription work done by for the transcribers on the Pecha tool web Interface.
  • To respond to the concern from Mr. David Newman regarding transcript cleanup and translation practice done for the srt files by his team using Gemini and provision of Openpecha tools for translation, Mr. Ganga Gyatso mentioned that he will be discussing with his colleague for new ideas after sharing his positive thought of using third party large language models for translation.
  • Mr. Ganga Gyatso also added that our team had completed transcription of16.41 hours of audio data on the basis of base model of STT. He added that 10 hours of fine-tuned Garchen Rinpoche STT model will be used for next batch of transcription. Also, mentioned that it would make easy manual transcription for our experienced transcribers.
  • Mr. Lhakpa Wangyal concluded importance of adhering to the PRD goal i.e. to build a custom Garchen Rinpoche Speech-to-Text model to transcribe Garchen Rinpoche’s teachings into timestamped transcripts (srt/WebVTT format). He added that after achieving the prime goal, its crucial to look for translation and further process.

Action Items

  • @DavidYesheNyima [Will be giving more audio data of Garchen Rinpoche] - Due: [ASAP]
  • @Ganga_Gyatso [will be update about the site for viewing the edited transcripts/srt files for the corresponding files] - Due: [second next coming week]
  • @Ganga_Gyatso [will be will be discussing with his colleague for pecha tool or other solutions regarding transcript cleanup and translation] - Due: [whenever it is possible]

Decisions Made

  1. Need more training data to achieve lower percentage of character error rate.
  2. To continue the transcription work using 10 hours of fine-tuned Garchen Rinpoche STT model.