# The demo recordings **Saturday repair café: the rain plan** is a scripted, AI-voiced planning call between three fictional volunteers. Rain forces their Saturday repair café indoors, but the hall isn't free until 10:30. They work through opening time, bike bookings, table space and who needs to call whom. - [Full Cassini file](repair-cafe.opus): about two minutes, three speakers. - [Short excerpt](repair-cafe-excerpt.opus): the first scene, about 42 seconds. - [Multitrack source recording](repair-cafe.multitrack.mkv): three separate, synchronized, lossless audio tracks, labelled Maya, Jonah and Tess. - [Authored script](repair-cafe.script.json) and [production record](repair-cafe.production.json). The public `.opus` files are produced by **Cassini itself** from the multitrack recording. Their transcripts are its unedited speech-recognition output. The script and external timing estimates are not supplied to Cassini. ## Generating the voices [Eleven v3 Text to Dialogue](https://elevenlabs.io/docs/overview/capabilities/text-to-dialogue) generated three continuous scenes using Alice as Maya, George as Jonah, and Lily as Tess. Continuous dialogue preserves the model's phrasing across turns. These are standard preset voices representing fictional people. Four short acknowledgments were generated in separate takes with their surrounding conversation as context. Only the acknowledgment audio was retained. They are placed over the principal speaker's continuing phrase at a slightly lower level: Tess's “Right,” Maya's “Yeah,” Jonah's “Mm-hmm,” and Tess's “Okay.” The overlaps exist in the audio, before any transcription or rendering. Main turns were separated at measured quiet boundaries into one synchronized track per person. Each acknowledgment went onto its own speaker's track. The three tracks sum back to the authored mix, within PCM quantization. This uses the original, unmixed TTS takes; no source-separation model is involved. The downloadable Matroska file contains three mono 24 kHz FLAC streams. Scribe v2 recognition was used only to locate edits in the source takes and check their spoken content. Long gaps between turns were shortened to varied 280–380 ms pauses. Within-turn hesitations remain. The opening of each scene is preserved, and acknowledgment clips retain quiet handles to avoid cutting consonants. Eleven v3's returned character timestamps were not used: they led the speech by up to two seconds in these takes. ## Processing with Cassini Cassini reads the three participant-labelled streams, mixes the listening copy, recognizes each speaker separately, measures cross-track attribution, and bounds word ends against that speaker's audio. The recipe uses its bundled **Parakeet TDT 0.6B v3 fp32** recognizer through sherpa-onnx on CPU. Automatic speaker-name hints are disabled for this example; optional summary generation is off. Speaker identity comes from the source stream metadata. Processing this source exposed a Cassini windowing bug: a VAD span contained an acknowledgment, a long quiet gap and a later sentence. Fixed decoding windows cut through that sentence and lost part of it. The included [Cassini patch](https://github.com/codemyriad/cassini-format/blob/main/site/scripts/cassini-vad-silence.patch) splits long VAD spans at sustained acoustic silence before applying the usual windowing to continuous speech. The production record identifies the base revision and patch, so this is a reviewable pipeline change. Cassini then packs the listening copy and recognized words into each `.opus` file. The site copies those exact bytes, without retagging or editing the transcript. [Full transcript JSON](repair-cafe.words.json) and [excerpt JSON](repair-cafe-excerpt.words.json) are decoded copies of their embedded word items. The production record includes output hashes and the recognizer/attribution provenance carried by the files. Recognition and word boundaries are model estimates, not manual ground truth. The script is available for comparison. The recognizer repeats “until” once in the opening sentence; that small error is retained in both files. Cassini's `provenance.wordTimings.endsBoundedByAudio` records its acoustic word-end check; it does not promise perfect recognition or sample-accurate word starts. ## The transcript interface The site mounts Cassini's actual `MeetingView` component, including its turn reconstruction, inline interjections, individual word hover/seek targets and playback highlighting. Both speakers' words can highlight during an overlap. Word times come from the file, with no timing interpolation in the player. Upstream PR #229 unintentionally removed word interactions. The restoration is tracked in [D-734](https://linear.app/code-myriad/issue/D-734/restore-word-level-transcript-highlighting-and-seeking-in-the-cassini). This site uses that same Cassini implementation; its integration patch does not add a separate word renderer or highlighting algorithm. ## Rebuilding The scripts in `site/scripts/` require Python with `requests` and `numpy`, plus `ffmpeg` and `ffprobe`. Voice generation and the editing reference use an ElevenLabs key with text-to-speech and speech-to-text permissions. Normal site builds make no AI API calls. ```bash python3 site/scripts/synthesize-demo.py --cache /tmp/repair-cafe-tts --generate python3 site/scripts/align-demo.py --cache /tmp/repair-cafe-tts python3 site/scripts/produce-demo.py --cache /tmp/repair-cafe-tts --out /tmp/repair-cafe-tracks ``` Set `ELEVENLABS_API_KEY` in the environment before generation. A matching cache avoids repeat API calls; changed inputs require a fresh cache. Review new takes because the model's seed is best effort. The alignment step rejects unexpected omissions or substitutions before the tracks are assembled. Build Cassini from the revision in the production record in an isolated checkout, with the included VAD patch applied. Then run the native pipeline: ```bash site/scripts/process-demo-cassini.sh /path/to/cassini \ /tmp/repair-cafe-tracks/repair-cafe.multitrack.mkv /tmp/repair-cafe.opus ``` For the excerpt, trim all three streams to the first scene before processing and packing them with the same command. `publish-demo.py` copies reviewed native outputs into the site and exports their embedded words and provenance; its `--help` lists the required source and processing records. Finally run `node site/scripts/gen-demo.mjs` to refresh the facts quoted by the pages. ## Earlier fixtures and licensing The original [Lantern Festival notes](lantern-festival.README.md) and recordings remain available for existing links. That older example used Kokoro-82M and interpolated word timings. `elements.opus` powers the separate song karaoke demo. The script, metadata and demo fixtures follow the repository's [licensing terms](../../../LICENSE.md). Voice models are not included. Cassini's component and the transcription patch are AGPL-3.0; the independent format readers and producer remain CC0.