Real-time audio transcription, automated editorial insight generation, and immediate multi-channel distribution — engineered specifically for the fast-paced interview and live moderation demands of CTV (Campus Media at the University of Applied Sciences St. Pölten / USTP).

In fast-paced editorial rooms, symposium discussions, and live interviews, editorial desks and moderators cannot afford cognitive distraction. They need precise transcripts, rapid extraction of key statements, and immediate social publishing drafts without manual transcription bottlenecks.

The CTV Audio Intelligence Engine addresses these constraints by pairing ultra-low-latency speech-to-text (Deepgram EU) with background semantic reasoning and editorial context steering.

The Editorial & Moderation Challenge

Live dialogue formats — such as expert panels, symposium discussions, campus interviews, and editorial roundtables — face three simultaneous operational bottlenecks:

1. Information Velocity: Rapid speaker statements and thematic pivots occur in seconds. Post-event transcriptions arrive too late for live follow-up questions or moderation steering.
2. Nomenclature & Proper Noun Accuracy: Academic titles, Austrian regional entities, Austrian Standard German phrasing, and guest names are frequently misspelled by generic speech engines.
3. Multi-Channel Distribution Pressure: Content desks must publish social media hooks, key quote cards, and summaries while the discussion is still ongoing.

┌────────────────────────────────────────────────────────────────────────┐
│             CTV Spoken Audio & Live Interview Stream                   │
│    (Studio Microphone · Stream Input · Panel Discussion & Debate)      │
└───────────────────────────────────┬────────────────────────────────────┘
                                    │ WebSockets (Deepgram EU)
                                    ▼
┌────────────────────────────────────────────────────────────────────────┐
│                 CTV Audio Intelligence Engine                          │
│       [ Custom Vocabulary: Names, Technical Terms, Acronyms ]          │
│       [ Session Context: Agenda, Topic Briefing, Guest Profiles ]      │
└─────────────┬─────────────────────┼─────────────────────┬──────────────┘
              │                     │                     │
              ▼                     ▼                     ▼
    ┌──────────────────┐  ┌──────────────────┐  ┌──────────────────┐
    │  Live Insights   │  │     Verbatim     │  │   Social Media   │
    │  (Sense-Making)  │  │ (Wortprotokoll)  │  │ (Live to Post)   │
    ├──────────────────┤  ├──────────────────┤  ├──────────────────┤
    │ Structured topic │  │ Clean, verbatim  │  │ Ready-to-publish │
    │ cards, entities, │  │ speech-to-text;  │  │ LinkedIn / X     │
    │ & action items   │  │ speaker diarized;│  │ teasers & quote  │
    │ directly to pool │  │ zero AI noise    │  │ hooks in seconds │
    └──────────────────┘  └──────────────────┘  └──────────────────┘

The Three Specialized Audio Modes

The engine provides three purpose-built operational modes selectable directly in the input bar:

1. 💡 Live Insights (sense-making)

UI Label: „Strukturiert Wissen aus dem Pool“

• Operational Mechanics: As the speaker talks, the audio stream is partitioned into semantic chunks (triggered dynamically upon reaching 50 words or detecting a 2-second pause in speech). A background serverless route analyzes each chunk alongside the active session history.
• Derived Artifacts:
  • Core Topic & Synthesis: Concise summary of the stated argument.
  • Key Entities: Identified organizations, projects, legislation, or technical methodologies.
  • Action Items & Questions: Follow-up angles and strategic inquiries for the moderator.
• Pool Ingestion: Extracted cards are dispatched directly into the session's Knowledge Pool as structured cards without cluttering the active chat window or disrupting the conversation flow.

2. 📄 Verbatim Protocol (verbatim)

UI Label: „Reines Transkript extrahieren“

• Operational Mechanics: Bypasses downstream LLM processing completely. Focuses strictly on high-precision speech-to-text with optional Speaker Identification (Diarization).
• Use Cases:
  • Legally sound interview records and verbatim documentation.
  • Rapid extraction of verbatim citations for editorial reporting and teleprompters.
  • Archival documentation of academic symposiums and press conferences.
• Efficiency: Operating in pure verbatim mode costs 1 Credit / Minute and eliminates unwanted AI hallucinations or rephrasings.

3. 📱 Social Media Drafts (social-media)

UI Label: „Live zu fertigen Posts“

• Operational Mechanics: Targets the digital editorial desk. While an expert, panelist, or guest speaks, the LLM pipeline synthesizes the main argument into publication-ready social media drafts.
• Derived Artifacts:
  • Headline hooks optimized for professional engagement (LinkedIn, X).
  • Direct pull-quotes with verified attribution.
  • Curated thematic hashtags matching the event or topic.
• Impact: Cuts the latency between a live spoken statement and its digital distribution to under 30 seconds.

Editorial Context & Custom Vocabulary

A core limitation of general-purpose voice software is phonetic distortion of niche terms, regional Austrian phrasing, and unfamiliar guest names. The CTV engine resolves this through dual context injection:

1. Custom Vocabulary (audioCustomVocabulary)

Located in the Live Audio Settings dialog, editors can configure a comma-separated list of critical terminology prior to recording:

• Guest Names: e.g. „Univ.-Prof. Dr. Hannes Raffaseder, Alois Frotschnig, Silvia Göhring“
• Institutions & Media: e.g. „USTP, Campus TV, Fachhochschule St. Pölten, ZID, c-tv“
• Technical & Regional Terms: e.g. „Neuroergonomie, Minto-Pyramide, Traisen, Niederösterreich“

These keywords are injected directly into the Deepgram WebSocket handshake, biasing the acoustic model toward phonetic accuracy.

2. Session Context & Editorial Briefing (audioSessionContext)

Moderators and editorial leads can paste the full discussion rundown, guest biographies, and thematic background into the session briefing context:

Session: CTV Editorial Roundtable #42
Topic: The Future of Media Production & Generative Interfaces
Moderator: CTV Editorial Team
Guests: Lead AI Researchers, Digital Media Ethics Specialists
Guiding Questions:
- How does generative AI alter editorial workflows and digital publishing?
- What are the copyright and EU AI Act implications for campus media?

This context is injected into every background LLM chunk call (/api/reactor/audio/sense-making). As a result, the generated Live Insights and Social Media posts recognize context, acronyms, and historical references that were not explicitly re-explained by the speaker.

Editorial & Moderation Workflow

A typical production cycle using the CTV Live Audio tool follows a three-stage sequence:

Step 1: Pre-Session Setup (3 Minutes)

1. Open the Live Audio Settings via the microphone menu or the quick gear icon.
2. Enter the expected guest names and specialist terminology into Custom Vocabulary.
3. Paste the interview guide or discussion agenda into Session Context.
4. Enable Speaker Identification if two or more microphones feed into the audio channel.

Step 2: Live Recording & Real-Time Moderation Monitoring

1. Select the required mode:
   • Live Insights for panel discussions, interviews, and editorial brainstorming.
   • Verbatim for formal lectures, addresses, and press conferences.
   • Social Media for breaking summaries and real-time social reporting.
2. Click Start Listening (🎙️). The real-time console expands at the top of the interface, displaying live incoming text alongside live background chunk deductions.
3. The moderation desk sees extracted insights drop into the right sidebar Pool in real time, serving as instant prompts for follow-up questions.

Step 3: Post-Session Synthesis & Multi-Export

1. Stop recording when the session concludes. The entire transcript is safely held in local and session memory.
2. Click Export to package the transcript into an Executive Memo, Typst Strategic Dossier, or raw text format.
3. Push selected insights to the canvas or send them to team members via a passwordless, read-only Share Snapshot Link.

Security, Privacy & EU Compliance

Because campus media organizations handle student, staff, and public guest data, audio processing must adhere to strict regulatory frameworks:

• EU Data Residency: All audio streams route exclusively to Deepgram's European Union endpoint (api.eu.deepgram.com).
• Zero-Retention Guarantee: Raw audio streams are processed in-memory as transient WebSockets. Neither the Brainstorm Reactor nor Deepgram stores the raw audio recording on disk. Audio payloads are purged immediately upon conversion to text.
• Role-Based Access Gate: Live transcription tokens are gated via /api/reactor/audio/token. Access is restricted to authenticated system administrators and verified members of the USTP / CTV organization (org_id: ustp).
• EU AI Act & GDPR Art. 6 Alignment: The system processes text objectively. Biometric voice profiling, emotion recognition, and persistent voiceprint storage are strictly excluded by design.