Speech Transcription Q&A¶

Transcribe audio speech, chunk the text, and write it to a knowledge base for later summaries, retrieval, and Q&A.

Prepare your resources¶

Prepare a meeting recording, a knowledge base, and an output volume. Use familiar content to check names, numbers, and key statements.

This template prepares searchable transcripts; Q&A takes place afterward. Compare transcripts with the recording, especially noisy passages and overlapping speech.

Use the template¶

On the workflow creation page, start from a template, search for Speech Transcription Q&A, and load it onto the canvas.

Complete the required fields in the parameter form.

Setting	Purpose
Input audio	Audio files or a volume containing them
Target knowledge base	Select an existing knowledge base or create one
Catalog save location	Location for the ZIP archive of chunked transcripts

Save the parameters, then start a manual run. The system saves the current workflow before submitting execution.

Complete processing flow¶

Arrows show execution order; configured bindings supply each step’s input.

Speech Transcription Q&A

Configure transcription and chunking¶

Setting	Default and use
Input audio	Select files or a volume and confirm scope
Language	`auto`; choose `zh` or `en` when the main language is known
Noise reduction	`auto`; alternatives are `off` and `ffmpeg_afftdn`; compare on a noisy sample
Chunk size	512 characters
Overlap	50; at least 0 and less than chunk size
Three-level index	Enabled, section size 5
Knowledge base and destination	Receive indexes and transcript files respectively

Chunk settings apply to text, not audio duration. Larger chunks do not repair transcription errors; verify transcription before adjusting retrieval context.

Check transcripts, indexes, and files¶

Confirm selected files and counts.
Compare parser documents and nonempty text with the original, checking names, numbers, and complete statements.
Inspect chunk context and source metadata.
Check index writes; three-level row counts differ from file counts.
Open the transcript ZIP at the destination and inspect records and source relationships.
Test retrieval using a question answered in the meeting recording.

The template writes chunked transcript documents to ZIP, saves that file, and registers sources. This differs from the document knowledge base template, which saves records before chunking.

Example: organize a meeting recording¶

Select a meeting recording, keep automatic language detection and default chunking, choose destinations, and run. Compare a key passage with its transcript, then retrieve the same topic. Adjust overlap or chunk size only after confirming transcription accuracy.

Common problems¶

Symptom	Check and action
Names or numbers are incorrect	Compare with speech and check language and noise
No text is returned	Check for recognizable speech and inspect parser output
Retrieved context is incomplete	Verify the transcript, then inspect chunks and overlap
Indexing succeeds but saving fails	Inspect later writing/saving status before repeating existing writes
Only one file is visible for multiple sources	Inspect all `file_ids` from document writing, not only the primary `file_id`

Last updated on Sep 15, 2026