Save and open projects as a single portable .hyperaudio file

[scarsellifi]

Summary

Add Save project / Open project as a portable .hyperaudio file.

A .hyperaudio file is a renamed ZIP containing:

• hyperaudio.json
• transcript.html
• optional transcript.original.json
• captions.vtt
• media/<original-file>

Principles

• Save, not export: original media is preserved byte-for-byte.
• Non-destructive editing: redactions and gap removal travel as data.
• Captions may intentionally diverge from transcript.
• One open project at a time in OPFS; filesystem is the project library.
• No API keys, credentials, or app preferences in the file.

Implementation

Add js/hyperaudio-save.js:
format → container → OPFS → editor bridge → UI.

Use JSZip for v1, whitelist-read on import, and OPFS for autosave.

Acceptance criteria

• Save downloads a .hyperaudio file.
• Open restores media, transcript, captions, options.
• Manual captions are not regenerated.
• Redacted words round-trip.
• Local autosave resumes the current project.
• Import treats the file as untrusted input.
• Remote (URL) media is embedded in the save whenever the server allows the
  download (CORS — Enable CORS on lab.hyperaud.io so the default media supports export and URL transcription #370); otherwise the project is saved as a declared
  media.kind: "link", and stays fully editable.
• When a project's media is unavailable on open (unreachable link URL, or a
  missing media entry), the editor offers to re-attach a local copy — the
  next save is self-contained again.

The full v1.1 format specification and a worked example are in the first
comment below.

Related

Supersedes the format sketched in #362 (the File System Access API remains a
possible future enhancement on Chromium); covers #378 via the provenance
block; with "the filesystem is the library", #363 reduces to saving the file
into a synced folder; replaces the localStorage flows of #169 / #182 / #239.

[scarsellifi]

The .hyperaudio file format — full specification (v1.1)

To the reader (human or LLM): this document specifies a file format. Rules
marked MUST / MUST NOT are binding for writers and readers; SHOULD
is a strong recommendation; MAY is optional behaviour.

---

1. What a .hyperaudio file is

A .hyperaudio file is the working save of a Hyperaudio Lite Editor
project: media + word-synchronized transcript + captions + project settings,
in a single portable file.

Technically it is a renamed ZIP archive (like .epub, .docx, .sketch),
designed to be read and written by:

• Hyperaudio Lite Editor (in the browser)
• future native apps (desktop / mobile)
• any tool that can open a zip and parse JSON

1.1 SAVE, not EXPORT — the founding principle

A .hyperaudio file captures the state of the work; it does not produce a
finished result. Three rules follow:

1. The container MUST NOT contain derived or rendered artifacts: no
   edited/cut media, no captions burned into frames, no re-timed transcript,
   no word-level karaoke VTT. Those are what the editor's export functions
   are for.
2. The media inside the container is the original, byte for byte. Saving
   never re-encodes anything.
3. Everything that is non-reconstructible working state goes in: the text
   with in-progress redactions, the current captions (which may have been
   hand-edited), the original machine transcription, the project settings.

In short: a .hyperaudio is a working zip that preserves the edits in
full — editing is non-destructive: decisions (redacted words, silences to
skip) travel as data (struck, options) on top of the intact original
media, and are applied at playback and at export time. Only rendered export
products stay out.

1.2 Format identity

Property	Value
Extension	.hyperaudio
MIME type	application/vnd.hyperaudio+zip
Container	ZIP (standard PKZIP)
Text encoding	UTF-8, always
Current version	1.1

---

2. Container structure

project.hyperaudio  (ZIP)
├── mimetype                   ← first entry, NOT compressed (see 2.1)
├── hyperaudio.json            ← THE file: version + media + options + texts + transcript (§ 3)
├── transcript.html            ← the transcript in the editor's native format, for compatibility (§ 4)
├── transcript.original.json   ← optional: the original machine transcription, immutable (§ 5)
├── captions.vtt               ← the current captions, WebVTT (§ 6)
└── media/
    └── <original filename>    ← the original media, STORE entry (§ 7)

Rules:

• hyperaudio.json MUST be present. It is the single source of truth for
  the working state.
• transcript.html and captions.vtt SHOULD be present (a conforming
  writer always writes them; a reader must tolerate their absence).
• transcript.original.json is optional: present only when the project
  was born from an automatic transcription (or from an import whose origin
  was kept).
• media/ contains at most one file. It may be absent when media.kind
  is "link" (§ 7.2.1).
• Readers MUST ignore unknown files in the container (this allows
  non-breaking future additions).

2.1 The mimetype entry (EPUB/ODF convention)

The first entry of the zip is a file named mimetype, containing exactly the
string:

application/vnd.hyperaudio+zip

with no trailing newline, stored without compression (STORE method) and
with no extra fields. This puts the MIME type bytes at a fixed offset of
38 in the physical file: a native app can recognise a .hyperaudio by
reading the first ~80 bytes, without even opening the zip.

• Writers MUST write this entry first, uncompressed.
• Readers SHOULD verify it, but MUST tolerate its absence (files
  produced by generic tools that re-zip the content remain valid).

---

3. hyperaudio.json — field-by-field reference

One JSON.parse and you have the whole project except the media. A complete,
valid example is in the collapsed section at the end of this comment.

3.1 Root

Field	Type	Required	Description
format	string	✔	Always "hyperaudio". Sanity check for readers.
formatVersion	string	✔	"major.minor", e.g. "1.0". Rules in § 8.
generator	object	✔	Who wrote the file: { "name": string, "version": string }. Diagnostic only.
created	string	✔	ISO 8601 UTC — when the project was first created.
modified	string	✔	ISO 8601 UTC — last save.
media	object	✔	Media descriptor (§ 3.2).
options	object	✔	Project settings (§ 3.3).
texts	object	✔	Key textual metadata (§ 3.4).
provenance	object	–	Origin of the transcription (§ 3.5).
transcript	object	✔	The complete working transcript (§ 3.6).

3.2 media

"media": {
  "kind": "original",
  "path": "media/intervista-maria.mp4",
  "url": null,
  "filename": "intervista-maria.mp4",
  "mimeType": "video/mp4",
  "durationSeconds": 62.5,
  "sizeBytes": 48211337
}

Field	Type	Description
kind	string enum	The media "formula": "original" (1.0) or "link" (1.1, § 7.2.1). Reserved value: "audio-m4a" (§ 7.2). Behaviour on an unknown kind: § 7.3.
path	string | null	Path inside the zip (with kind: "original" / "audio-m4a"). null with "link". Security constraints in § 10.
url	string | null	URL of the media on the web (only with kind: "link"). Otherwise null.
filename	string	The original filename chosen by the user, preserved.
mimeType	string	MIME type of the media (video/mp4, audio/mpeg, …).
durationSeconds	number	Duration in seconds (float). Informational; also used for reconciliation (§ 7.3).
sizeBytes	number	Size of the media file in bytes. Informational; also used for reconciliation (§ 7.3).

3.3 options — project settings

"options": {
  "gapRemoval": { "enabled": true, "thresholdMs": 500, "bufferMs": 100 },
  "captions": { "updateFromTranscript": false },
  "view": { "showSpeakers": true, "showTimecodes": false }
}

Field	Type	Description
gapRemoval.enabled	bool	Silence skipping active in preview.
gapRemoval.thresholdMs	int	Minimum pause (ms) for a silence to be skipped.
gapRemoval.bufferMs	int	Margin (ms) kept at the edges of a skipped silence.
captions.updateFromTranscript	bool	The flag that governs divergence (§ 6.1). true = captions are derived; the editor regenerates them on every transcript edit. false = captions are hand-curated: the editor MUST NOT overwrite them.
view.*	bool	Project display preferences.

Readers MUST ignore unknown keys inside options; writers may add keys
with a minor version bump.

⚠️ options MUST NOT ever contain: API keys, tokens, application
preferences (preferred transcription engine, etc.). Those are app
preferences, not project settings — a .hyperaudio file gets shared; a key
inside it is a guaranteed leak.

3.4 texts — key metadata

"texts": {
  "title": "Intervista con Maria — formato di salvataggio",
  "language": "it",
  "summary": "Maria e Piero presentano il nuovo formato .hyperaudio.",
  "topics": ["hyperaudio", "formato file", "salvataggio"]
}

Clean data (strings and arrays, never HTML). title is also the basis of the
filename suggested at download. language is BCP-47 ("it", "en-GB").
summary and topics may come from the transcription engine or from the
user; they may be an empty string / empty array.

3.5 provenance (optional)

"provenance": {
  "engine": "deepgram",
  "model": "nova-3",
  "transcribedAt": "2026-07-10T08:55:00Z",
  "originalTranscript": "transcript.original.json"
}

Field	Type	Description
engine	string	Engine that produced the transcription (deepgram, whisper, parakeet-local, …).
model	string	Model used, if known.
transcribedAt	string	ISO 8601 UTC of the original transcription.
originalTranscript	string	Path of the file holding the original machine transcription (§ 5), if kept.

Records who/what produced the original transcription. Zero cost today,
valuable tomorrow (e.g. deciding whether to re-transcribe with a better
model). Absent when the transcript was pasted/imported without keeping the
origin.

3.6 transcript — the working transcript

"transcript": {
  "words": [
    { "start": 0.32, "end": 0.84, "text": "Benvenuti" },
    { "start": 0.84, "end": 1.02, "text": "ehm", "struck": true }
  ],
  "paragraphs": [
    { "speaker": "Maria", "start": 0.32, "end": 6.5 }
  ]
}

words[] — every word, in temporal order:

Field	Type	Description
start	number	Start in seconds (float).
end	number	End in seconds (float).
text	string	The word, without its trailing space.
space	bool	true = a word boundary (space) follows the word. false = fragment glued to the next one (e.g. hyphenated words split by the engine). Default when absent: true.
struck	bool	true = redacted word (struck out): excluded from playback and from every export, but text and timings remain. Default when absent: false. Do not serialize defaults (smaller files, backward compatibility with pre-existing JSON).

paragraphs[] — the paragraph/turn structure:

Field	Type	Description
speaker	string | null	Speaker name (without square brackets), or null if unattributed.
start / end	number	Temporal extent of the paragraph in seconds. Words belong to the paragraph that temporally contains them.

Rules:

• Times are seconds (float) throughout the JSON. (The editor's DOM uses
  integer milliseconds — data-m/data-d; the conversion is
  ms = round(s × 1000); § 12.)
• Redactions (struck) are working state and travel in the save. They
  are reversible in the editor; they become final only in an export.
• In v1, redaction is word-level only (struck), complemented by
  gap-removal in options. Time-range redaction may arrive as a future
  additive field redactions: [{start, end}] (minor bump, e.g. v1.1) if a
  real need emerges.

---

4. transcript.html — the compatibility copy

The transcript in the editor's native format: HTML with one <span> per
word.

<article><section>
  <p>
    <span class="speaker" data-m="320" data-d="0">[Maria] </span>
    <span data-m="320" data-d="520">Benvenuti </span>
    <span data-m="840" data-d="180" style="text-decoration: line-through;">ehm </span>
  </p>
</section></article>

Conventions of the HTML format (the editor's current ones):

• data-m = start in milliseconds (integer); data-d = duration in
  milliseconds.
• The trailing space inside the span encodes the word boundary
  (equivalent to space: true in the JSON).
• Speakers are dedicated spans: class="speaker", data-d="0", text
  [Name] .
• Redactions are style="text-decoration: line-through" on the span
  (equivalent to struck: true).

Role:

1. Compatibility: existing HTML-based flows (hyperaudio-lite, legacy
   storage) consume it without converters.
2. Inspectability: opening the zip, a browser displays the transcript
   directly.
3. Safety net: if the JSON round-trip ever had a bug, the editor's data
   is still there.

Anti-divergence rule: the source of truth is
hyperaudio.json.transcript. The writer MUST generate transcript.html
from the same state in the same save (the two files are consistent by
construction).

Fallback rules: the reader MUST load from the JSON; using
transcript.html as a source is allowed only as recovery, when the JSON
is missing or unreadable. In that case the reader:

• MUST sanitize the HTML against the allowlist in § 10 before any DOM
  insertion;
• SHOULD warn the user that the file is not fully conforming and was
  recovered from the HTML;
• SHOULD regenerate the complete JSON at the next save, bringing the
  project back to full conformance.

---

5. transcript.original.json — the origin (optional, immutable)

The transcription as it came out of the engine, before any human
intervention. It is non-reconstructible working state: the moment the user
edits, the original would be gone — and re-transcribing costs time/money and
never yields the exact same result.

Schema: identical to the transcript object of hyperaudio.json
(§ 3.6) — { "words": [...], "paragraphs": [...] }, times in seconds. A
diff between origin and working copy is then a direct comparison of equal
structures.

Rules:

• The writer writes it once, at transcription time (or at import, if the
  origin is kept), and MUST NOT ever modify it afterwards. It is the
  conceptual twin of the media: it enters at the start and is never touched.
• It MUST NOT contain struck: redaction is the user's work, absent from
  the origin by definition.
• Speakers are those assigned by the engine (e.g. "Speaker 0"), not the
  names the user will assign later.
• The reader MUST tolerate its absence and MUST NOT load it as the
  working transcript, except on an explicit user action (e.g. "restore
  original").
• It is referenced by provenance.originalTranscript (§ 3.5); if the file
  exists but the field is missing (or vice versa), the reader trusts the
  file.

What it enables (future features, none required in v1): a "what did I
change" diff view, restoring a word/region to the original, re-alignment,
engine quality metrics.

The origin always travels in the file: stripping it from the save would
only make sense if it were public and retrievable online — it is not, so the
file is the only place it lives. There is no "save without origin" option in
v1.

⚠️ Privacy note: the origin contains the pre-redaction text — words the
user struck out or deleted in the working copy are still there. Within a save
this creates no new exposure class (the container already includes the
original media, which contains all the audio), but any future sharing
feature that strips the file (e.g. "share without media") MUST consider
stripping this file too.

---

6. captions.vtt — the current captions

The project's sentence-level captions, in standard WebVTT.

6.1 ⚠️ Transcript and captions MUST be allowed to diverge

This is the most important conceptual point of the format, and the reason
captions are a file of their own rather than a derivative recomputed on open.

Transcript and captions are two representations of the same speech with two
different purposes:

• The transcript is the display and editing view: faithful word by
  word, with word-level timing, redactions included. It is the document you
  work on — the equivalent of the timeline.
• The captions are the reading view: text segmented into cues sized to
  be read while watching — short lines (~37 characters), minimum
  duration, sustainable reading speed (~17 characters/second), line breaks
  placed where the sentence breathes.

A good subtitle is not a faithful transcription: it may omit a
hesitation, compact a repetition, break a line at a different point, nudge a
cue by a few tenths of a second to give it time to be read. Whoever curates
subtitles in caption mode produces irreplaceable user data, not a
derivative.

The contract is governed by options.captions.updateFromTranscript:

Value	Meaning	Reader obligation
true	The captions are derived: the editor regenerates them on every transcript edit. The saved VTT is the latest generation (included because it makes the file readable by a player with no segmentation logic).	May regenerate them freely.
false	The captions are hand-curated and legitimately diverge from the transcript.	MUST NOT regenerate or "fix" them to re-align them to the transcript. Touch them only on an explicit user action.

Corollary for every reader (including automated tools and LLMs processing the
format): never "repair" a file by re-aligning captions and transcript. If
they diverge with updateFromTranscript: false, the divergence is
intentional.

6.2 What this file is NOT

• It is not the word-level karaoke VTT: that is an artifact always
  deterministically derivable from the transcript and MUST NOT be in the
  container.
• It is not the burned-in caption of an export: that is produced from the
  re-timed transcript at export time and is not working state.

6.3 Multilingual captions (roadmap)

captions.vtt at the root is and will remain the primary track, forever
— including in a multilingual future. Additional tracks will arrive as an
additive minor bump: captions/<lang>.vtt files declared in
hyperaudio.json. v1 readers will keep working by finding the primary; the
extra tracks are simply unknown files to ignore (§ 2).

---

7. media/ — the project media

7.1 v1 rules (kind: "original")

• The file enters the container byte for byte, never re-encoded. The
  original name is preserved (media/<filename>).
• In the zip, the media entry SHOULD use the STORE method (no
  compression): media formats are already compressed; deflate would gain ~0%
  while burning CPU on files hundreds of MB large. JSON/HTML/VTT are
  compressed normally (deflate).
• One media file per project.

7.2 Future formulas (roadmap, not v1)

The media.kind field is the enum that makes these evolutions a non-breaking
minor bump:

kind	Content	Status
"original"	The original file in media/	v1 — the only initial formula
"audio-m4a"	Only the audio track, re-encoded M4A/AAC: a compact save for projects where the video weighs GBs but the work is on the text. An explicit user choice at save time, never automatic.	future
"link"	No file in the container: url points to media on the web. The file is not self-contained and the editor states so openly.	v1.1 (§ 7.2.1)

7.2.1 kind: "link" (since 1.1)

• url MUST be an http(s) URL; path is null; the container has no
  media/ entry.
• The file is not self-contained, and readers state so openly on open.
• Writers SHOULD prefer embedding over linking: attempt to download the
  remote media and save it as kind: "original" (whether this works is the
  server's call — CORS); fall back to a link save only with the user informed.
• Readers open a link project by playing the URL directly (playback does not
  require CORS). If the URL is unreachable, § 7.3 applies — degraded mode
  with reconciliation is the recommended behaviour.

7.3 Media unavailable: degraded mode and reconciliation

When the media is unusable — unknown kind (written by a newer version),
kind: "link" with an unreachable URL, or a tampered container — the reader
MUST NOT load the project as if nothing happened. Two conforming
behaviours:

1. Rejection with a clear message (the default; this is what Hyperaudio
   Lite Editor v1 does).
2. Degraded mode (MAY): load the text only (transcript/captions),
   explicitly stating that the media is unavailable.

In degraded mode the reader MAY offer reconciliation: ask the user to
provide the media file and re-attach it to the project. Verification criteria
in v1: durationSeconds, sizeBytes, filename (heuristic); once a
sha256 checksum lands in the descriptor in v1.x, reconciliation becomes
certain. Re-attaching rewrites the media descriptor (back to
kind: "original") at the next save.

---

8. Versioning

formatVersion is a "major.minor" string.

• minor bump (1.0 → 1.1): backward-compatible additions — new fields,
  new files in the container, new values where the reader has a safe fallback
  behaviour. Readers MUST ignore unknown fields and files.
• major bump (1.x → 2.0): breaking changes. A reader that encounters
  a major version above its own MUST reject the file with a clear message
  ("this project requires a newer version of the editor"), without attempting
  a partial load.
• Writers always write the most recent version they know and MUST NOT
  silently rewrite a file to a lower version.

Ignore-unknown + reject-major is the pair of rules that lets the format
evolve for years without breaking native apps born later.

Version history: 1.0 initial; 1.1 adds media.kind: "link" (§ 7.2.1).

Documented exception: unknown media.kind → the project is not normally
loadable even within the same major; the behaviours of § 7.3 apply.

---

9. What must NEVER be in the container

1. API keys, tokens, credentials — of any service, in any field.
2. Application preferences (preferred engine, UI language, etc.) — only
   project settings belong here.
3. Derived artifacts: edited media, burned-in captions, karaoke VTT,
   re-timed transcript.
4. Data from other projects, or editor history.

---

10. Reader security (normative)

A .hyperaudio file comes from the outside: the reader treats it as
untrusted input.

10.1 Whitelist-read: never generic extraction

The reader MUST NOT iterate the zip entries extracting them or writing
them to paths taken from the file. It reads only entries with known
names: mimetype, hyperaudio.json, transcript.html,
transcript.original.json, captions.vtt, plus the media indicated by
media.path. Every other entry is ignored (§ 2). This neutralises path
traversal (../, absolute paths) by design: no path from the file is ever
used as a write destination.

10.2 Constraints on media.path

• MUST match the pattern media/<filename> — a single segment after
  media/, no further /, no .., no absolute paths.
• MUST correspond to an existing entry in the zip.
• Only that entry is read as media; any other file in media/ is ignored.

On violation, the reader MUST treat the media as unavailable (§ 7.3).

10.3 Size limits (anti zip-bomb)

Before decompressing, the reader SHOULD enforce reasonable ceilings on
the textual entries (indicative: 50 MB each for hyperaudio.json,
transcript.html, transcript.original.json, captions.vtt — a one-hour
transcription is on the order of hundreds of KB). Entries above the ceiling →
the file is treated as non-conforming.

10.4 JSON validation

The reader MUST validate types and ranges before use: numeric times
finite and non-negative, start ≤ end, well-formed formatVersion,
media.kind among the handled values. Malformed or invalid JSON =
"unreadable JSON" → the recovery path from transcript.html applies (§ 4).

10.5 Safe DOM construction

• Primary path (from the JSON): the reader MUST build the DOM
  programmatically, using textContent for all text — never innerHTML on
  data coming from the file.
• Recovery path (from transcript.html): the reader MUST sanitize
  against an allowlist before any DOM insertion — allowed elements:
  article, section, p, span; allowed attributes: data-m, data-d,
  class (value speaker only), style (only
  text-decoration: line-through). Everything else — scripts, on*
  handlers, iframes, links, images, other styles — is removed.

10.6 Duplicate entries

With whitelist-read, duplicate entries are not an attack vector (a single
entry per name is read — the one the zip library deterministically exposes).
Readers SHOULD nevertheless not depend on entry order, mimetype
excepted.

---

11. Conformance checklist

A conforming writer:

• writes mimetype as the first entry, STORE, exact content
• writes hyperaudio.json with all required fields, times in seconds,
  defaults (space: true, struck: false) not serialized
• writes transcript.html and captions.vtt consistent with the JSON of
  the same save
• writes transcript.original.json once (at transcription time) and
  never modifies it again
• writes the media byte for byte, STORE entry, name preserved — or, for
  a link save (§ 7.2.1), no media entry and an http(s) url
• never writes keys/credentials/derived artifacts

A conforming reader:

• checks format === "hyperaudio" and applies the version rules (§ 8)
• reads only entries with known names; validates media.path and the
  size limits (§ 10)
• validates JSON types and ranges before use (§ 10.4)
• loads the transcript from hyperaudio.json, building the DOM with
  textContent; uses transcript.html only as recovery, sanitized
  (§ 4, § 10.5)
• honours captions.updateFromTranscript: false (never
  regenerate/re-align)
• never loads transcript.original.json as the working transcript
  (explicit user action only)
• on unavailable media: rejects with a clear message, or degrades while
  stating it (§ 7.3)
• ignores unknown fields and files
• tolerates the absence of mimetype, transcript.html,
  transcript.original.json, captions.vtt

---

12. Invariants and known traps

Invariant	Why it matters
JSON in seconds (float), DOM/HTML in milliseconds (integers)	The off-by-1000 error is the classic fatal bug of this domain. Conversion: ms = round(s × 1000); maximum loss 0.5 ms, irrelevant.
Trailing space in the HTML span = space: true in the JSON	The word boundary is data, not formatting. Losing it glues words together.
struck travels in the save, is applied at export	Redactions are reversible as long as you are in the working file.
Captions ≠ transcript (§ 6.1)	Never "repair" the divergence: with updateFromTranscript: false it is intentional.
Origin ≠ working copy (§ 5)	transcript.original.json is immutable and pre-redaction: never load it in place of the working transcript, never "update" it.
The file is untrusted input (§ 10)	Whitelist-read, validation, DOM via textContent, sanitization of the HTML fallback.
UTF-8 everywhere	Media filenames included (the zip's UTF-8 flag).
One media file, original name	No ambiguity, no index to maintain.

Worked example — a complete, valid container

Zipping these entries (with mimetype first and uncompressed) produces a
valid .hyperaudio file (a 1.0 file — still valid under 1.1, which is a
purely additive minor bump). The example demonstrates the two key contracts:
caption divergence (§ 6.1 — the transcript contains the redacted hesitation
"ehm", the hand-curated captions omit it) and the immutable origin (§ 5 — the
engine produced "hyper" + "audio" as two lowercase words; the user merged
them into "Hyperaudio,").

mimetype

application/vnd.hyperaudio+zip

hyperaudio.json

{
  "format": "hyperaudio",
  "formatVersion": "1.0",
  "generator": { "name": "hyperaudio-lite-editor", "version": "0.8.2" },
  "created": "2026-07-10T09:00:00Z",
  "modified": "2026-07-10T11:30:00Z",

  "media": {
    "kind": "original",
    "path": "media/intervista-maria.mp4",
    "url": null,
    "filename": "intervista-maria.mp4",
    "mimeType": "video/mp4",
    "durationSeconds": 62.5,
    "sizeBytes": 48211337
  },

  "options": {
    "gapRemoval": { "enabled": true, "thresholdMs": 500, "bufferMs": 100 },
    "captions": { "updateFromTranscript": false },
    "view": { "showSpeakers": true, "showTimecodes": false }
  },

  "texts": {
    "title": "Intervista con Maria — formato di salvataggio",
    "language": "it",
    "summary": "Maria e Piero presentano il nuovo formato di salvataggio .hyperaudio.",
    "topics": ["hyperaudio", "formato file", "salvataggio"]
  },

  "provenance": {
    "engine": "deepgram",
    "model": "nova-3",
    "transcribedAt": "2026-07-10T08:55:00Z",
    "originalTranscript": "transcript.original.json"
  },

  "transcript": {
    "words": [
      { "start": 0.32, "end": 0.84, "text": "Benvenuti" },
      { "start": 0.84, "end": 1.02, "text": "ehm", "struck": true },
      { "start": 1.1, "end": 1.3, "text": "a" },
      { "start": 1.3, "end": 2.1, "text": "Hyperaudio," },
      { "start": 2.3, "end": 2.55, "text": "il" },
      { "start": 2.55, "end": 3.1, "text": "modo" },
      { "start": 3.15, "end": 3.4, "text": "più" },
      { "start": 3.45, "end": 4.05, "text": "semplice" },
      { "start": 4.1, "end": 4.3, "text": "di" },
      { "start": 4.35, "end": 5.0, "text": "montare" },
      { "start": 5.05, "end": 5.3, "text": "un" },
      { "start": 5.35, "end": 6.1, "text": "video" },
      { "start": 6.15, "end": 6.5, "text": "trascritto." },
      { "start": 7.8, "end": 8.2, "text": "Grazie" },
      { "start": 8.25, "end": 8.7, "text": "Maria," },
      { "start": 8.9, "end": 9.3, "text": "oggi" },
      { "start": 9.35, "end": 9.9, "text": "parliamo" },
      { "start": 9.95, "end": 10.15, "text": "del" },
      { "start": 10.2, "end": 10.75, "text": "formato" },
      { "start": 10.8, "end": 10.95, "text": "di" },
      { "start": 11.0, "end": 11.85, "text": "salvataggio." }
    ],
    "paragraphs": [
      { "speaker": "Maria", "start": 0.32, "end": 6.5 },
      { "speaker": "Piero", "start": 7.8, "end": 11.85 }
    ]
  }
}

transcript.html

<!-- Compatibility copy: the transcript in the editor's native format.
     Source of truth: hyperaudio.json → transcript (this file is generated from the same save).
     data-m = start in milliseconds, data-d = duration in milliseconds.
     The trailing space inside each span encodes the word boundary.
     style="text-decoration: line-through" = redacted word (struck in the JSON). -->
<article><section>
  <p>
    <span class="speaker" data-m="320" data-d="0">[Maria] </span>
    <span data-m="320" data-d="520">Benvenuti </span>
    <span data-m="840" data-d="180" style="text-decoration: line-through;">ehm </span>
    <span data-m="1100" data-d="200">a </span>
    <span data-m="1300" data-d="800">Hyperaudio, </span>
    <span data-m="2300" data-d="250">il </span>
    <span data-m="2550" data-d="550">modo </span>
    <span data-m="3150" data-d="250">più </span>
    <span data-m="3450" data-d="600">semplice </span>
    <span data-m="4100" data-d="200">di </span>
    <span data-m="4350" data-d="650">montare </span>
    <span data-m="5050" data-d="250">un </span>
    <span data-m="5350" data-d="750">video </span>
    <span data-m="6150" data-d="350">trascritto. </span>
  </p>
  <p>
    <span class="speaker" data-m="7800" data-d="0">[Piero] </span>
    <span data-m="7800" data-d="400">Grazie </span>
    <span data-m="8250" data-d="450">Maria, </span>
    <span data-m="8900" data-d="400">oggi </span>
    <span data-m="9350" data-d="550">parliamo </span>
    <span data-m="9950" data-d="200">del </span>
    <span data-m="10200" data-d="550">formato </span>
    <span data-m="10800" data-d="150">di </span>
    <span data-m="11000" data-d="850">salvataggio. </span>
  </p>
</section></article>

transcript.original.json

{
  "words": [
    { "start": 0.32, "end": 0.84, "text": "benvenuti" },
    { "start": 0.84, "end": 1.02, "text": "ehm" },
    { "start": 1.1, "end": 1.3, "text": "a" },
    { "start": 1.3, "end": 1.7, "text": "hyper" },
    { "start": 1.75, "end": 2.1, "text": "audio" },
    { "start": 2.3, "end": 2.55, "text": "il" },
    { "start": 2.55, "end": 3.1, "text": "modo" },
    { "start": 3.15, "end": 3.4, "text": "più" },
    { "start": 3.45, "end": 4.05, "text": "semplice" },
    { "start": 4.1, "end": 4.3, "text": "di" },
    { "start": 4.35, "end": 5.0, "text": "montare" },
    { "start": 5.05, "end": 5.3, "text": "un" },
    { "start": 5.35, "end": 6.1, "text": "video" },
    { "start": 6.15, "end": 6.5, "text": "trascritto" },
    { "start": 7.8, "end": 8.2, "text": "grazie" },
    { "start": 8.25, "end": 8.7, "text": "maria" },
    { "start": 8.9, "end": 9.3, "text": "oggi" },
    { "start": 9.35, "end": 9.9, "text": "parliamo" },
    { "start": 9.95, "end": 10.15, "text": "del" },
    { "start": 10.2, "end": 10.75, "text": "formato" },
    { "start": 10.8, "end": 10.95, "text": "di" },
    { "start": 11.0, "end": 11.85, "text": "salvataggio" }
  ],
  "paragraphs": [
    { "speaker": "Speaker 0", "start": 0.32, "end": 6.5 },
    { "speaker": "Speaker 1", "start": 7.8, "end": 11.85 }
  ]
}

captions.vtt

WEBVTT

NOTE
HAND-CURATED captions (options.captions.updateFromTranscript = false).
They legitimately diverge from the transcript: the hesitation "ehm" is
omitted, line breaks are chosen for reading, and the second cue extends
past the end of its last word (6.5s -> 6.7s) to leave time to read it.
A conforming reader MUST NOT regenerate them or re-align them to the
transcript.

00:00:00.320 --> 00:00:03.100
Benvenuti a Hyperaudio,
il modo più semplice

00:00:03.150 --> 00:00:06.700
di montare un video trascritto.

00:00:07.800 --> 00:00:11.850
Grazie Maria, oggi parliamo
del formato di salvataggio.

media/

In a real file this folder contains the original media, byte for byte, with
its filename preserved (here: intervista-maria.mp4, a STORE entry). Omitted
from this example.