Audio-to-News-Article Agent
An AI agent that turns a recording into a publishable news article
- 2024
- Year
- Hugging Face
- Hosted on
- Ses → Metin → Haber
- Pipeline
Overview
For a reporter, the most time-consuming part of the job is not writing the story — it is transcribing the recording. Transcribing a half-hour interview routinely takes over an hour, and all of that time is non-creative labour.
This agent was built to remove that step. An uploaded audio file first passes through a speech recognition model, producing time-stamped raw text. A language model then shapes that text into news format: headline, standfirst, lede and body. Quotes from the speaker are marked separately, so the editor can see which sentences are direct quotations.
The system deliberately does not "auto-publish". The output is always a draft; the final call stays with the editor. The goal is not to replace the reporter but to lift the transcription and first-draft burden off them.
Highlights
- Time-stamped transcription from an audio file
- Automatic structuring into news format (headline, standfirst, body)
- Direct quotes marked separately
- Runs in the browser on Hugging Face Spaces
Objectives
- Take transcription time out of human hands
- Produce a publishable first draft from raw speech
- Keep editorial control with a person
Solution
- Chaining speech recognition and the language model into one pipeline
- An output template that enforces news structure
- Teaching the model to separate quotes from commentary
- A Streamlit interface that needs no technical knowledge
Approach
- 01Accuracy testing with real interview recordings
- 02Refining the output template together with an editor
- 03Deploying on Hugging Face Spaces
Outcomes
- Transcription became a matter of minutes rather than hours
- Reporters can spend most of their time on verification and writing
- Draft structure became consistent across publications
Challenges
- Speech recognition accuracy on noisy recordings
- Preventing the model from blending quotes with its own sentences
- Accuracy of proper nouns and terminology in Turkish recordings
Selected Work
Similar projects
Voice Transcript App
A Streamlit application that generates automatic transcripts from uploaded audio files and lets you download the result as a Word document.
PDF Q&A Agent (ChatPDF)
A RAG-based assistant that analyses uploaded PDFs and answers user questions grounded in the document's own content.
YouTube Transcript App
A Streamlit application that generates a transcript from a YouTube link and delivers it as a Word document.