How Multilingual Speech-to-Text Is Accelerating Lecture Captioning for an Online University

The Challenge

Captioning had depended on a mix of outsourced vendors and staff edits, and turnaround slipped every time faculty uploaded a large batch of new content.

The existing approach made it difficult to:

• Caption new lectures before students needed them, especially at the start of each term.
• Recognize technical terms, formulas, author names and course-specific jargon correctly.
• Handle instructors with diverse accents and lectures that switch between English and Spanish.
• Offer translated subtitles for international learners without a separate manual workflow.
• Track caption quality and coverage across the whole library for accessibility reporting.

From a technical perspective, lecture recordings vary widely. Some are studio-produced, while others come from laptop microphones, classroom capture systems or live webinars with audience questions. The pipeline had to detect language changes within a recording, keep caption timing aligned with slides and speech, meet caption formatting standards for line length and reading speed, and integrate with the university's video platform and LMS without adding work for faculty.

The Solution

InnovationM designed an event-driven pipeline that starts when a video lands in the media platform and ends with captions ready for review in the same place. The team built an evaluation set from recorded lectures across several departments and used it to compare speech engines and measure every tuning change. The solution included:

• Audio preprocessing: voice activity detection, noise reduction and loudness normalization to clean up classroom and home-office recordings.
• Multilingual speech recognition: a speech-to-text layer with automatic language identification that handles English, Spanish and mixed-language segments within one lecture.
• Course-aware vocabulary: custom vocabularies built automatically from syllabi, slide decks and reading lists so subject terms and names are recognized correctly.
• Caption formatting: segmentation into readable caption lines with timing, line-length and reading-speed rules, exported as WebVTT and SRT files.
• Machine translation for subtitles: translated subtitle tracks for selected languages, flagged for review by bilingual staff where courses require it.
• Review and reporting workspace: an editor where accessibility staff correct low-confidence segments, with corrections fed back into vocabulary tuning and a dashboard showing caption coverage and quality across the library, built with InnovationM's data engineering practice.

The key focus was not simply generating captions faster, but making captions accurate enough that staff review became a quick check rather than a rewrite.

The Impact

Captioning moved from a backlog managed by vendors and spreadsheets to an automated step in the publishing workflow, with people focusing on review and quality. The solution supported:

• Shorter time from lecture upload to published captions.
• Lower word error rate on course-specific vocabulary after tuning.
• Higher share of the media library with reviewed captions.
• Reduced captioning cost per hour of content.
• Searchable transcripts that students use to revisit specific parts of a lecture.

By bringing together multilingual recognition, course-aware vocabulary, standards-based caption formatting and a review loop, the online university can publish accessible lectures on schedule and report on coverage with confidence.