The recording.
The words.
One file.

An ordinary audio file that carries its own transcript.

Cassini keeps audio, word timestamps and speaker labels together in a single .opus file. Send it, save it, build on it. The words travel with the recording.

Try a file Build with Cassini

Standard Ogg Opus. Open specification. No sidecar to lose.

Listen & follow along.opus

Saturday repair café: the rain plan

3 speakers · 1:53 · Click a word to seek

Reading Saturday repair café: the rain plan…

File details Reading…

read the audio file

walk the Ogg pages, find OpusTags

reassemble CASSINI_PAYLOAD_000..N

gunzip, parse, check SHA-256

resolve the default transcript

Audio + transcript, from the same file Download 0.86 MB

Three fictional volunteers, voiced with ElevenLabs v3 on separate speaker tracks. Cassini transcribed the recording and produced this file. About this example

One .opus fileAudio and text together

Word-level timingRead, search and jump to speech

Ordinary playbackWorks in players that support Opus

Open to implementCC0 schemas and reader code

01 / Why keep them together?

A recording is more useful
with the words attached.

A transcript should survive the trip from the tool that made it to the person who needs it.

Share the whole conversation

A meeting recording, interview or podcast can carry who said what. Copy one file and its transcript comes along.

Choose how to use it

Listen in an Opus player. Read and seek in a Cassini viewer. The format also works for music with synchronized lyrics.

Keep it useful for later

Original words and timings can sit beside display text and additional transcripts, with provenance recording what produced them.

02 / Inside the file

Audio underneath. Context built in.

Cassini adds metadata where an audio file already keeps its title and artist: the OpusTags header.

meeting.opusone file

Same extension. Same audio. More to work with.

Transcript

Words you can read, search and seek

Word timestamps and speaker labels live in the audio file’s comment header. A Cassini reader turns them into a transcript you can navigate. Multiple transcripts can share the same recording.

A word from the example file

{
  "speaker": "spk_maya_a95db5b0ac159e4384ff55ef",
  "startMs": 60,
  "endMs": 459,
  "text": "So,"
}

Look at the actual tags and decoded manifestFor a closer look

The manifest and each transcript are UTF-8 JSON, gzipped, base64url-encoded and split across numbered comments. These values come from the downloadable example.

repair-cafe.opus 43 comments

CASSINI_ATTRIBUTION_MODE=annotate

CASSINI_ATTRIBUTION_RAN=true

CASSINI_ATTRIBUTION_WORDS_DROPPED=0

CASSINI_ATTRIBUTION_WORDS_FLAGGED=0

CASSINI_ATTRIBUTION_WORDS_MEASURED=28

CASSINI_AUDIO_CHANNELS=1

CASSINI_AUDIO_DURATION_MS=113113

CASSINI_AUDIO_MATCH_POLICY=exact-opus-audio-v1

Payload values are shortened for display.

Decoded manifest

{
  "kind": "cassini-portable-meeting",
  "version": 1,
  "profile": "ogg-opus",
  "meeting": {
    "id": "mtg_8d28b76a269092cd722ccaee50953dce9de30021c00146f33170c58ab50c0c80",
    "title": "Saturday repair café: the rain plan",
    "createdAtUtc": "2026-09-07T16:48:00.244576383Z",
    "processedAtUtc": "2026-09-07T16:48:00Z",
    "durationMs": 113113
  },
  "audio": {
    "container": "ogg",
    "codec": "opus",
    "sampleRate": 48000,
    "channels": 1,
    "sampleCount": 5429448,
    "durationMs": 113113
  },
  "integrity": {
    "matchPolicy": "exact-opus-audio-v1",
    "opusAudioSha256": "8d28b76a269092cd722ccaee50953dce9de30021c00146f33170c58ab50c0c80",
    "sampleRate": 48000,
    "channels": 1,
    "sampleCount": 5429448,
    "durationMs": 113113
  },
  "speakers": [
    {
      "id": "spk_maya_a95db5b0ac159e4384ff55ef",
      "label": "Maya"
    },
    {
      "id": "spk_tess_7134aac53eddcc48c4c90667",
      "label": "Tess"
    },
    {
      "id": "spk_jonah_b2d60d92ca43aa51a0c91ba9",
      "label": "Jonah"
    }
  ],
  "transcripts": [
    {
      "id": "raw-asr",
      "default": true,
      "format": "cassini.words.v1",
      "wordCount": 331,
      "createdAtUtc": "2026-09-07T16:48:00Z",
      "payloadRef": {
        "prefix": "CASSINI_TX_RAW_ASR_PAYLOAD_",
        "chunkCount": 2,
        "sha256": "5a51f4c677c5637050a9019ac1c5824eb93891329423e967f941a51e53829505",
        "rawBytes": 31850,
        "gzipBytes": 4427,
        "mime": "application/vnd.cassini.transcript-words+json",
        "encoding": "base64url+gzip+utf8json"
      }
    }
  ],
  "provenance": {
    "speechToText": {
      "raw-asr": {
        "backend": "sherpa-onnx",
        "model": "parakeet-tdt-0.6b-v3",
        "device": "cpu",
        "hints": {
          "termCount": 3,
          "decodingMethod": "greedy_search",
          "applied": false,
          "reason": "disabled by configuration (CASSINI_STT_HINTS_DISABLED)"
        }
      }
    },
    "attribution": {
      "ran": true,
      "mode": "annotate",
      "wordsMeasured": 28,
      "wordsFlagged": 0,
      "wordsDropped": 0
    },
    "wordTimings": {
      "endsBoundedByAudio": true
    }
  }
}

Learn how to extract and verify the payload →

In this example, the compressed manifest and transcript total 5.2 KB in a 0.86 MB file. These sizes exclude base64 encoding and tag overhead.

03 / Make something with it

Start with a file. Build from there.

Copy a reader, use the Python producer, or implement the specification in your own stack.

Extract & display

Read a file

Get the transcript out in Python or JavaScript. Learn how to decode, verify and handle incomplete metadata.

Guide + working readers Pack & preserve

Write a file

Combine an Opus recording with a word-timed transcript using a complete producer in standard-library Python.

Guide + working producer Implement & validate

Use the specification

Find the v1 contract, transcript format, audio digest, JSON Schemas and conformance suite in one place.

Reference + test vectors

Working with a coding assistant?

Give it the complete specification in one text file

Before you build

A few useful boundaries.

Version 1 is published and used by gocassini, the reference implementation. The ecosystem is still small. Read the project status →

Does Cassini create the transcript?

The format stores a transcript you already have. Recording and speech recognition happen in a producer such as gocassini. The standalone Python producer takes audio and timed words as input.

Do listeners need a Cassini app?

They need a player that supports Ogg Opus to hear the audio. A Cassini reader is needed to display the embedded words and speaker labels. The browser reader is one example.

What happens if the audio is edited?

Audio edits can leave the embedded transcript out of date. A reader that checks the audio digest can detect a mismatch and label the transcript as stale. Software may also strip metadata, so check files after editing or converting them.

Do the digests prove authenticity?

No. They check that data matches, including whether a transcript refers to the same audio. Anyone rewriting a file can recompute its hashes. Cassini does not provide signatures or proof of who said something.

Can I implement it in my own software?

Yes. The schemas, test vectors and standalone readers and producer are CC0. The specification text is CC BY 4.0. The transcript interface shown here comes from Cassini and is AGPL-3.0, as is the gocassini application. Start with the reference →

The recording.The words.One file.