Navigate

Quick find

Find what you need

Esc
Approved narrationCandidate v6 audio withheld

ERAIS / V6 film room

The V6 film room.
Approved voice. Open evidence.

The approved 185.11-second founder-voice anchor is published with its measured captions and transcript. The five V6 demonstration narration clips are candidates pending founder approval, so their audio is withheld; their scripts and event provenance stay visible in the arena.

Film identity

Approved master185.11 s · 24 kHz · founder-voice anchor
Web encode185.108 s · AAC · 2139 KiB
Captions38 cues · VTT/SRT pinned
V6 candidate clips5 · audio withheld
Visual masters3 separate masters
EvidenceV6 discovery arena

Only the approved V5 voice anchor is published as audio. The V6 clips await founder approval.

Approved narration / captions / transcript

Listen to the approved anchor.

This is the approved V5 founder-voice master, 185.11 seconds, with its measured captions. The player carries the approved audio only; the visual masters below are separate artifacts, not a synchronized render.

Transcript — 11 segments
  1. Field note, twenty-second of November, twenty twenty-six. This is a long-form test of the ERAIS native presentation voice: one founder capsule, one continuous session, no donor runtime in the inference path.
  2. The pipeline has four visible stages. Text is normalised, phonemised and mapped to a fixed token table. A four-layer Zipformer text encoder turns those tokens into a conditioned sequence.
  3. A five-stack acoustic velocity network integrates four anchor-flow steps, each step predicting a clean mel trajectory from the current state, the text condition and a six-second speech reference.
  4. The final mel is decoded by a native forty-eight kilohertz Vocos graph: eight ConvNeXt blocks, two upsampling stages, two ISTFT heads, a sinc resampler and a Linkwitz crossover.
  5. The voice itself is a reference-free capsule: the conditioning tensors learned from the approved take, stored as ordinary float weights. Inference does not need the source recording.
  6. For the first time, the same acoustic and vocoder parameters execute through ERAIS-owned graph code. The imported weights remain explicitly attributed.
  7. The important engineering question is no longer whether the voice can run without the donor. It is which operations deserve native weight-space programs, which should remain deterministic DSP, and where the current fracture boundaries waste time or constrain expression.
  8. The measurements in this run separate tokenisation, text encoding, four velocity calls, and vocoder decode. They also count attention score maps, feed-forward projections, convolution branches and memory high-water marks.
  9. That matters because a two-minute waveform is not one model call. The qualified context is roughly forty-two seconds including the reference, and full attention grows quadratically. Long-form narration therefore uses the largest context-safe segments and one continuous voice state, not a single impossible forward pass.
  10. The next upgrades are targeted. Replace repeated dense projections with composable weight-space banks. Make attention, feed-forward and convolution schedules parametric and measurable. Cache immutable conditioning and reference structure. Increase segment size where the context allows.
  11. Then compare every change against this same founder voice on naturalness, cadence, pronunciation and long-form consistency.

Existing visual masters / separate audio

What the film can show.

These are the existing visual masters, reused unchanged. Each has its own audio and page; none is presented as a synchronized 185-second render of the approved narration.

V6 candidate narration / approval boundary

The V6 clips are not published yet.

The five V6 demonstration clips are candidates pending founder approval. Only their metadata, caption counts and hashes are listed; no candidate voice audio is present on this page or in the public media directory.

ClipDurationCaption cuesStatus
V6 master (candidate)
00_master
260.8 s52audio withheld
Autonomous discovery
01_discovery
54.5 s11audio withheld
Generalization and consolidation
02_generalization
45.1 s9audio withheld
Neural Doom Cyborg
03_cyborg
46.7 s9audio withheld
Native voice
04_voice
33.7 s8audio withheld

The five v6 narration clips are candidates pending founder approval. No candidate audio is published; only metadata, captions hashes and scripts are listed.

Evidence / links

Follow the claims.

The film points to the source-pinned evidence rather than replacing it. The approved anchor, the V6 discovery arena, the live players and the execution observatory remain separately inspectable.

Release receipt: reports/agi_flagship_v6/agent_b_public_release.json

Claim boundary

  • The approved V5 master is the founder-voice anchor; it is not a V6 demonstration-script film.
  • The v6 demonstration narration is pending founder approval and is withheld.
  • The visual masters are separate audio/video artifacts, not a synchronized render.
  • Model-based voice metrics are proxies, not a blinded human MOS study.
  • Donor-origin acoustic/vocoder weights remain; WSP coverage is partial.