STT0073: LLM-Based Correction of Inference Transcriptions Using Claude AI

[jim-gyas]

Description:

Develop a process that improves the quality of inference transcriptions for audio files using Claude AI by aligning them with a verified transferred text. The transferred text is known for its spelling accuracy, while the inference transcription captures the audio's context more effectively. The goal is to create a corrected transcription that combines the strengths of both. The process will involve generating inference transcriptions, feeding both inference transcriptions and transferred text into Claude AI, and using its output to replace the transferred text in the original dataset.

Resources:

1) Inference Transcription and Transferred Text : news-catalog

Completion Criteria:

• Claude AI processes both the inference transcriptions and transferred text to output corrected text.
• Corrected text maintains the context and meaning of the inference transcription while ensuring spelling accuracy based on the transferred text.
• A new CSV file is generated where the Transferred Text column is replaced with the Corrected Text column, retaining all other original columns.

Implementation:

Subtasks:

• [x] Generate Inference Transcription for audio files using your inference function from the stt-split-audio tool.
• [x] Run the transfer function to obtain Transferred Text, which is considered accurate for spelling.
• [x] Develop a prompt for Claude AI that correct the spelling errors in the Inference Transcription by referencing the Transferred Text.
• [ ] Feed both Inference Transcription and Transferred Text (matched by filename) into Claude AI. Claude outputs the Corrected Text.
• [ ] Replace the Transferred Text column with the Corrected Text, Retain all other columns from the original CSV and then Save the resulting file.

[jim-gyas]

API account has insufficient credit to access the Claude API.

[jim-gyas]

@kaldan007 and @gangagyatso4364, could you please review my card?

[gangagyatso4364]

add the step to make cer comparison between all three kinds of transcript generated at last.

[gangagyatso4364]

looks good go ahead.