The Problem with Generic TTS

Standard text-to-speech (Google Polly, Amazon Polly, even ElevenLabs) produces technically correct but emotionally flat audio. Listeners subconsciously recognize "robot voice" and disengage. For a news channel, this kills retention.

What We Use Instead

Our pipeline uses two voice systems in sequence:

• Fish Speech (primary) — Open-source TTS model with natural prosody. We run it locally for primary narration.
• Custom RVC Voice Cloning — Retrained Voice Cloning model trained on our own voice data. This produces indistinguishable-from-human narration in our voice.
• gTTS (fallback) — When both primary systems fail, Google's free TTS ensures the pipeline never blocks on voice.

How RVC Cloning Works

RVC (Retrained Voice Cloning) is a voice conversion technique. It doesn't just synthesize speech — it takes a base voice output and converts it to sound like the target voice. The process:

1. Collect ~30 min clean voice samples (no background noise)
2. Train RVC model (RINDEX = 0, lr = 1e-4, batch = 8)
3. Export voice features (.pth file ~380MB per voice)
4. At inference: base model output → RVC conversion → final audio

Why It Matters

Brand identity. A news channel's voice IS the brand. Generic TTS is interchangeable. Our voice model makes The ViralDNA instantly recognizable — like any human presenter would be. At ~12 seconds of audio per sentence of processing time, it's slow. But quality wins.

The Fallback Chain

Voice is the slowest phase (~10 min per run). If inference hangs or audio comes out garbled, the pipeline tries gTTS and continues. The story ships with lower-quality voice rather than not shipping at all. We review voice failures separately and reprocess when needed.

One Production Cycle, Start to Finish

Here is exactly what happens when the pipeline runs. No hand-waving. Real timing data from production runs.

Phase 1 — Discovery (5 min)

The system queries 300+ news sources via Serper API and RSS feeds. Topics are clustered using semantic similarity. Duplicates are merged. What emerges is a ranked list of "story candidates" for the day.

Phase 2 — Weighting (2 min)

Each candidate is scored on relevance, timeliness, audience demand, and diversity (vs. yesterday's output). The top 50 stories get produced. Weighting ensures we never publish three stories about the same topic.

Phase 3 — Quality Gate (3 min)

Legal risk and compliance filter. Checks for unverified claims, sensitive topics, and potential defamation. Stories that fail the gate don't get produced — better to skip than risk a legal issue. This is non-negotiable for a Telugu news channel.

Phase 4 — Scripting (5 min)

Gemini Flash-latest generates a full journalistic script with RAG augmentation (it pulls context from search results, not just its training data). Output: ~400 words, structured like a real news segment. Hook → Context → Development → Close.

Phase 5 — Voice (10 min)

The slowest phase. Fish Speech generates base audio, then RVC converts it to our voice. ~12 seconds processing per sentence of script. This is why voice is the bottleneck. A fallback to gTTS exists but is rarely used.

Phase 6 — Thumbnail (2 min)

AI-generated thumbnail with large, centered text overlay. The pipeline uses reinforcement learning feedback from CTR data to improve click-through over time. Text must be bold and readable on mobile screens.

Phase 7 — Assembly (10 min)

FFmpeg assembles everything: background images with Ken Burns effect (slow zoom/pan), synchronized subtitles, voice track, branded intro/outro overlays, transitions. Output: one MP4 file, 1920x1080, ~3-5 minutes long.

Phase 8 — Forensic Audit (1 min)

Pre-ship quality check. Validates: audio present and synced, resolution correct, duration within bounds, no corruption. If it fails, the story is rejected. No garbage ships to YouTube.

Phase 9 — Upload (5 min)

YouTube Data API v3 uploads the video. Two outputs: main video (full length) and a Short (vertical cut, <60s). Metadata: title, description, tags, category, scheduling. Videos go live on a schedule (main: 1 hour after upload, Shorts: 30 min after).

Phase 10 — Post-Pipeline (5 min)

SEO article published to the blog, sitemap updated, community engagement posts scheduled, analytics tracking initialized. Each video generates a companion article for Google indexing. The cycle begins again for the next story.