Output Formats Explained (What They Are and Why You'd Use Them)¶
OpenScreech can produce several output formats. This guide explains each format in plain English and helps you pick the right one for your workflow.
These formats change how recognized text is represented or packaged. Enabling additional formats does not change the underlying recognition quality.
Quick Choosing Guide¶
- If you only need readable text: use
TXT. - If you need subtitles for most editors/players: use
SRT. - If you need web video subtitle compatibility: use
WebVTT. - If you need broadcast/enterprise caption workflows: use
TTMLorSMPTE-TT. - If you need model-transformed text (summary/cleanup): use
.llm.txt. - If you need to debug what was sent to the model: use
.llm-request.json. - If you need word timing data: use
.words.jsonl. - If you need a chunk-aware record of who spoke when: enable
transcript.speakers.txt.
Multitrack Output Scopes¶
Multitrack recording uses the selected formats for per-track exports, combined exports, or both. Combined files are at the session root; per-track files are in numbered tracks/ folders.
Each enabled scope includes transcript.speakers.txt using track names, independently of the ordinary recording diarisation toggle. Overlapping utterances are labelled in text. Combined captions split at utterance activity boundaries and can include multiple labelled lines in one cue; the optional .words.jsonl retains individual word timestamps, track IDs, and track names.
Word timing is always required internally for merging, even when its sidecar export is off. Enabling multiple formats does not transcribe the audio again. The optional LLM processor runs on combined output; the separate per-track LLM option adds individual requests.
Transcript and Caption File Formats¶
TXT (transcript.txt)¶
What it is:
- Plain text transcript.
Why use it:
- Easy to read, search, copy, and edit.
- Works in every text editor.
Best for:
- Notes, documents, archives, simple post-processing.
SRT (transcript.srt)¶
What it is:
- Subtitle format with numbered cues and time ranges.
Why use it:
- Very widely supported in video tools and media players.
Best for:
- Editing subtitles in common video workflows.
WebVTT (transcript.vtt)¶
What it is:
- Web subtitle format for HTML5-style players.
Why use it:
- Better native compatibility in web-based playback workflows.
Best for:
- Website video players, browser-first media pipelines.
TTML (transcript.ttml)¶
What it is:
- XML-based timed text format used in some professional caption systems.
Why use it:
- Required by some enterprise and media delivery systems.
Best for:
- Advanced caption distribution pipelines.
SMPTE-TT (transcript.smpte-tt.ttml)¶
What it is:
- A broadcast-focused profile of TTML.
Why use it:
- Required by certain professional broadcast/media standards.
Best for:
- Broadcast-specific subtitle/caption workflows.
Word Timing Sidecar (transcript.words.jsonl)¶
What it is:
- Word-level timing data in JSONL (one JSON object per line).
Why use it:
- Enables fine-grained timing analysis and custom downstream tooling.
Best for:
- Advanced QA, custom subtitle tooling, analytics.
Speaker-Labelled Transcript (transcript.speakers.txt)¶
What it is:
- A separate transcript that keeps the recording's ASR chunks and labels speaker turns inside each chunk.
Why use it:
- Makes meeting narrative and turn-taking easier to follow without replacing
transcript.txt. - Explicitly shows overlapping speakers and unattributed speech instead of silently guessing.
Best for:
- Meetings, interviews, panels, and other multi-speaker recordings.
Notes:
- It reuses the same chunked ASR words as
transcript.txt; the diariser supplies only speaker time ranges. - Cue splitting is applied inside speaker turns, after chunk and speaker boundaries.
Meeting LLM Processor Input¶
Processor Input is an optional Meeting Recording setting, not another output file. It chooses the single recording representation sent to the selected LLM processor.
The default is Original Transcript. It sends the recognized transcript text without requiring any additional saved output and is the recommended choice for ordinary summaries, cleanup, and note generation. Existing meeting workflows continue to use this option automatically.
Alternative inputs are available only when their matching recording outputs are enabled:
| Processor Input | Requirement | Typical Reason to Use It |
|---|---|---|
Original Transcript |
None beyond live transcription | Normal summaries and transformations |
Transcript Text Output (.txt) |
Enable Transcript Text output | Include the saved recording header and chunk markers |
Speaker-Labelled Transcript (.speakers.txt) |
Enable speaker-labelled output and configure a diariser | Preserve speaker turns and overlap labels |
Word-Level Timing Sidecar (.words.jsonl) |
Enable the word-level timing sidecar | Give custom analysis exact word timing records |
OpenScreech sends only the selected representation. The processor's optional beginning and end tags wrap that text, which becomes the chat request's user message; the processor instructions remain a separate system message when Send system prompt is enabled.
OpenScreech does not automatically truncate oversized input. Long recordings—and especially word-level timing JSONL—can exceed a local model's context window, so choose the smallest representation that contains the information the prompt needs.
LLM Post-Processor Output Files¶
LLM Post-Processor Output (transcript.llm.txt)¶
What it is:
- Text returned after the completed transcript is sent to the selected LLM post-processor.
Why use it:
- Gives you a transformed version without overwriting the original transcript.
Best for:
- Summaries, polished notes, structured markdown outputs.
LLM Request Payload (transcript.llm-request.json)¶
What it is:
- The exact JSON request body OpenScreech sends using the selected LLM post-processor.
Why use it:
- Debugging prompt/model/provider behavior.
- Reproducibility when tuning prompts.
- Inspecting the rendered prompt and transcript content included in a request. Treat this file as sensitive.
Best for:
- Troubleshooting and model-output QA.
Paste Formats (Live Dictation Output)¶
These are profile settings, not saved subtitle files.
Paste Input Format¶
Plain Text: treat text as plain text.Markdown: treat text as markdown semantics before conversion.HTML: treat text as HTML semantics before conversion.
Why this matters:
- It tells OpenScreech what your output text means before paste conversion.
Paste Format¶
Plain Text: paste as plain text only.Rich Text: paste as styled rich text.HTML: paste as HTML-oriented clipboard content.Auto (Multi-format): write multiple clipboard formats and let the destination app choose.
Why this matters:
- Different apps prefer different clipboard formats.
Autooften works best across mixed apps, but plain text is most predictable.
HTTP API Response Formats¶
For HTTP transcription requests:
response_format=text: returns plain transcript text.response_format=json: returns JSON containing transcript text.
Choose based on your client:
- Humans/scripts reading raw output:
text - Apps expecting structured parsing:
json
Recommended Default Bundles¶
Everyday Notes¶
- Enable:
TXT - Optional:
.llm.txt - Disable: advanced subtitle formats unless needed
Video Editing (General)¶
- Enable:
TXT+SRT - Optional:
WebVTT
Web Publishing¶
- Enable:
TXT+WebVTT - Optional:
SRT
Broadcast/Enterprise Delivery¶
- Enable:
TXT+TTML+SMPTE-TT - Optional:
SRTfor internal review
Common Mistakes¶
- Enabling every output format by default and creating unnecessary clutter.
- Using only
.llm.txtand forgetting to keep the sourceTXTfor reference. - Expecting all video tools to prefer the same subtitle format.
Quick Tips¶
- Keep
TXTwhenever an LLM output is enabled. It gives you the recognized source needed to audit a summary or rewrite. - SRT, WebVTT, TTML, and SMPTE-TT package timing differently; enabling another caption format does not rerun or improve ASR.