Every newsroom says it is automated. Almost none of them can show you the moment their own pipeline refused to ship a bad take. This is that moment, from an Arabic geoeconomic bulletin that four stories a day assembles, voices, scores, and renders itself.
THE CHOKEPOINT (نقطة اختناق) is 2TInteractive's MSA news product, built on SceneWeave. It sits on 614 wired feeds from 423 publishers, scores what is worth a segment, re-scripts it as broadcast Arabic, diacritises it locally, voices it, builds the graphic plates, and renders a 16:9 and a 9:16 master from one narration stem. Episode 02 went out at 8:09 with four files: the Fed's October minutes, the IMO carbon-neutrality fight, an ADNOC tanker attack off Hormuz, and Waller pushing for a December hike.
Episode 01 was the buzz cut. We shipped it with known Arabic pronunciation roughness because the point was to prove the loop ran end to end. Episode 02 is where the loop had to start proving it could be trusted.
What the second episode got wrong
A listener flagged it before our QA did: inside Story 1, the same clause was being said twice.
Two defects, both real audio:
- A verbatim repeat. At 1:04.8 the voice delivered بدلاً من التوقف عند المستوى الحالي, then restarted the sentence from that same six-word clause. Three and a half seconds of duplicate.
- A garbled restart. At 2:11 the closing clause was attempted twice: the first attempt mangled, the second correct, so the ending was heard twice, the first one broken.
The script was clean. No sentence was duplicated in the copy. The duplication was generated inside a single text-to-speech inference.
The root cause was architectural, not cosmetic
Our speech preprocessor coalesces adjacent Arabic spans into one span so that a single loudness pass and one crossfade chain stitch a take cleanly. Story 1 contains no Latin-script tokens, so each half of the story collapsed into a single 35 to 43 second span, and that whole span went to XTTS-v2 in one call. XTTS tail-loops on generations that long: it repeats the final clause, or restarts a clause and garbles the first attempt.
Here is the part that matters more than the glitch. Our QA gate passed it. The gate tested word error rate and a loose word-count ratio. The defective take scored a WER of 0.200 and a ratio of 1.067. That ratio was the repeated clause, counted as four extra words. And it passed, because nothing in the pipeline asked the one question that mattered: did the voice say something it was not asked to say twice?
A gate that does not test for a defect is not a gate. It is a dashboard.
What we changed in the machine
- A real speech chunker. Takes are split at clause boundaries against a budget measured in base letters, not words, and not bytes, because Arabic diacritics inflate byte counts without adding audio. No single synthesis call now carries more than about eight seconds of speech. The stitching reuses the existing interior-silence trim, 50 ms quarter-sine crossfade, and single loudnorm pass.
- A repetition guard. A new check detects internal n-gram repeats, fuzzy repeats (one substituted token, because ASR spells the second pass slightly differently), script-anchored injected repeats, and gross word inflation or loss. It is deliberately conservative: ASR spelling variation must never trip it.
- A gate that fails the build. Repetition is now a hard failure. The build refuses to render a take that loops. The chunking policy is part of the cache identity, so changing the policy invalidates cached audio instead of silently reusing a bad take.
The numbers
| Segment | WER before | WER after |
|---|---|---|
| 01-a | 0.200 | 0.150 |
| 01-b | 0.147 | 0.040 |
| 02-a | 0.033 | 0.011 |
| 03-b | 0.192 | 0.138 |
| All 10 segments | 10/10 "passing" | 10/10 clean, 0 repetitions |
Verified on the delivered master, not just the stems: the previously defective passages at 0:57 to 1:15 and 2:13 to 2:32 now deliver each clause once.
Episodes
Episode 01, the pilot. Episode 02, the quality pass. Both stream on the OTT, self-hosted.
What automation has to mean
It is easy to automate the happy path. The test of a production system is whether it catches its own defects before an audience does. Episode 02 is the first time ours did the catching, and the fix landed in the pipeline rather than in a one-off file, so episode 03 inherits it.
Concretely, that is what we are selling: a wired newsroom that produces a bilingual-ready, dual-aspect bulletin on demand, and can show you the receipt when something goes wrong. 614 feeds. 423 publishers. One narration stem, two masters. Ten self-scored segments. A gate that fails on repetition.
What is next
MSA is the quality floor we are mastering now. The commercial second act is dialect: the same pipeline, routed to Gulf, Egyptian, and Levantine registers, with the same gate proving each one. If you are building a channel, a briefing product, or an internal newsroom and you want to see the machine rather than a slide about it, that is what SceneWeave does.
The same team ships sovereign apps, and takes on platform engineering and media production work. The episodes themselves live on our OTT.