Studio
Open a media file from Projects, or load one locally.
AI Data & Model Operations Dashboard
Dedomena Labs' workbench for multilingual speech, language, vision, model evaluation, call-center AI, media localization, conversational AI, and production delivery. The project and media selectors control the active operational context.
Core business
Speech & Voice AI
ASR evaluation, diarization, telephony transcription, dialect and domain testing, TTS evaluation, human gold data, and custom speech model workflows.
Media AI & Localization
Professional transcription, subtitles, dubbing, translation, live captions, media indexing, and human review for multilingual content.
Conversational AI
Call-center AI, voicebots, chatbots, multilingual conversation data, response evaluation, safety review, and domain adaptation.
Vision & Multimodal AI
Image and video labeling, scene and object annotation, OCR, video indexing, CV or VLM evaluation, and cross-modal quality review.
Operational coverage
Global Speech & Dubbing Intelligence
Current project / media
Media Convert & Resample
Video Indexing Lab
Transcript / dub snapshot
Workbench entry points
Load the 15-clip Swedish speech-labeling demo set to inspect annotation, metadata, QA and export workflows.
Open Speech / Audio labeling for manual segmentation, transcription correction, tags and speaker metadata.
Re-open a native Dedomena Labs annotation session for continued labeling or evaluation.
Domain / topic coverage (this session)
Platform capabilities
- Global speech data & ASR operations - Global Auto combines high-accuracy multilingual transcription with speaker diarization, then detects language per turn and code-switching. Remote WhisperX/pyannote remains optional.
- Transcript normalization & adjudication - cached punctuation, casing, Arabic orthography and proper-name cleanup with glossary/context; raw ASR is preserved.
- Versioned annotation sets - ASR, subtitle and dub groupings coexist; re-segmenting no longer destroys the previous cut.
- Language data & selective translation - source can be Auto detect or any typed language/code. Target-language speech can stay original while other languages receive separate subtitle translation and spoken dub adaptation.
- TTS / voice evaluation & synthesis - OpenAI standard TTS is available for previewed/locked actor voices. Source-voice dubbing uses the Dedomena Labs GPU clone endpoint with one stable reference per speaker. Azure remains optional for final validation.
- Realtime speech workflows - browser audio streams through Dedomena Labs' server-side WebSocket proxy so the OpenAI key never reaches the browser. Public capture requires HTTPS.
- Computer vision / video intelligence - scene-change + periodic sampling, cached transcript context and OpenAI vision produce timestamped shot/entity/topic/on-screen-text records.
Local VAD vs. server transcription
The built-in ✨ Auto-segment in Label Audio is still a local WebAudio energy-VAD. It is useful for manual labeling because it detects speech/silence without uploading anything, but it does not transcribe or identify speakers.
For production media, use Studio → ① Transcribe. Global Auto is now the default: it combines multilingual high-accuracy ASR with diarization, detects language per turn and editorial speaker roles, then caches the result before deriving ASR, subtitle or dub segment sets.
Speech / Audio Labeling
Load a clip, click ✨ Auto-segment to detect speech regions automatically, then click into any transcript box to type. Speech over 15s auto-splits.
Shortcuts: space play · S split · M merge · A auto-segment · J/K next/prev · Del remove
Source audio
Segments (0)
ASR Import / Advanced
Normal server transcription now lives in Studio and follows the active project/media automatically. Use this page only for an external GPU endpoint, Whisper/WhisperX JSON, SRT/VTT import, or other advanced/manual ingestion.
Transcription endpoint
not configured{audio, model, language} and expects Whisper-style {segments:[{start,end,text,speaker?}]} back.Or import an existing transcript
Pipeline - one raw transcript, many outputs
Transcribe (here)
Audio → time-stamped transcript via your ASR endpoint, or import existing Whisper/SRT output.
Post-edit & tag → Label Audio
Transcribers correct the machine output and add non-speech/markup tags. The 50-70% margin lever: verify, don't transcribe from scratch.
Translate → Translate / TTS QA
Per-segment translation with length-bounds validation for TTS fit.
Subtitle → Export SRT/VTT
Caption-ready export with the CPS/duration readability rule applied.
Dub → Dubbing / Side-by-Side
Source + target text side by side per segment, with per-row playback and synth hooks.
Live CC / Live Translation
Stream microphone or browser/tab audio through the Dedomena Labs server. The OpenAI key stays on the droplet. The active project automatically supplies source/target language and context.
Source / live captions
Translated captions
Notes
Realtime transcription is optimized for low-latency captions and does not provide speaker labels. For final delivery, run the recorded media through the diarized file-transcription pipeline afterward. For browser/tab capture your browser may require selecting “share audio”.
Dubbing / Side-by-Side
Source ↔ target per segment, with per-row playback, merge (same-speaker guarded), split-at-second, and synth hooks. Load the media in Label Audio (or here), and rows from a CSV or the Raw Processing tab.
Segments (0)
Translate / TTS QA
Per-segment translation review, mirroring the source → target workflow with a TTS "Now Playing" strip for reviewing synthesized voice output against the transcript.
Segments (0)
No speech segments yet - add some in Label Audio first, then come back here.
Label Images
Computer-vision labeling workspace: draw bounding boxes, assign classes and attributes, and export training/evaluation annotations to COCO JSON or YOLO txt.
Annotations (0)
No boxes yet - draw on the image above.
Label Video
Video / vision-model labeling workspace: temporal actions, frame-level gesture regions, object/track metadata and validation-ready annotations for training or model evaluation.
Video annotations (0)
No annotations yet.
Subtitle & Dub Studio
Applied media-delivery workspace for documentary subtitling and dubbing. Works on the segments from Label Audio - QC against broadcast readability rules, build the dub script with duration budgets, export to delivery formats.
Dub pipeline - Colab interop (parameters mirror your notebooks)
🎙 Voice Studio - TTS & cloning
💰 Job pricing - cost vs. Azure list
Cue list & dub script (0)
No segments yet - create them in Label Audio (or import an SRT above).
Author Prompts
Write prompt-completion pairs to the Dedomena Labs SFT schema - the 2023-style production workspace. RTL-aware, live word/char counts, per-item validation badge.
No items yet - click ＋ New item, or import a .jsonl of {"prompt","completion",...} objects.
Evaluate Models
Explore a broad global model catalog, build a benchmark shortlist, then run human rubric scoring and pairwise preference judgments across language-model, speech, vision, and multimodal outputs.
Explore model catalog Hugging Face Hub first layer
Benchmark shortlist
0 selectedExport ratings
QA & Agreement
Inter-annotator agreement and gold-set scoring - the numbers that go in every client deliverable. Compare two annotators' work, or score an annotator against a gold standard.
Transcription agreement (two session JSONs)
Eval rating agreement (two ratings JSONLs from the Evaluate Models tab)
Reports: per-dimension exact agreement %, mean absolute difference (1-5 scale), and Cohen's κ for pairwise A/B/Tie choices. These are the statistics quoted in the Eval Pilot deliverable.
Validation Rules
Production QA gate - run all rules before delivery. Rules with no applicable data show N/A rather than failing.
Rule thresholds
Constraint settings (transcription conventions)
File Metadata & Speaker Roster
Required file-level values and per-speaker labels, per section 4 of the transcription guidelines.
File-level values
Speaker roster
Tag Reference
Full non-speech, markup, and disfluency tag set consolidated from the en_US v3.0 transcription guidelines and the IFAP 16kHz addendum.
Non-speech noise tags - human vocal
| [breath] | Inhalation/exhalation between words, yawning |
| [cough] | Coughing, throat clearing, sneezing |
| [cry] | Crying / sobbing |
| [laugh] | Laughing, chuckling |
| [lipsmack] | Lipsmacks, tongue-clicks |
Non-speech / non-human noise tags
| [applause] | Clapping - placed exactly where it occurred (16kHz addendum) |
| [beep] | Beep replacing profanity or classified info |
| [bg-speech] | Background speech overlapping the foreground speaker |
| [click] | Machine or phone click |
| [dtmf] | Telephone keypad tone |
| [music] | ≥1s of music/singing with no foreground speech; on-hold music |
| [no-speech] | Pause / silence of ≥1s, even with faint foreground noise mixed in |
| [noise] | Any other miscellaneous noise (screaming, rain, punching, etc.) |
| [ring] | Telephone ring |
| [sta] | Continuous static |
Markup tags
| <initial></initial> | Wraps an initialism/acronym spoken letter-by-letter, e.g. <initial>TV</initial> |
| <lang:Foreign></lang:Foreign> | Wraps a word/phrase in an unidentified non-target language |
| <lang:X></lang:X> | X = any capitalized language name (Arabic, Korean, Spanish, English…) |
| <overlap></overlap> | Wraps speech that overlaps with another speaker |
Disfluency & other markup
| #eh / #mm / #öh | Filler words - prefixed with # |
| word~ | Stumbled / truncated word, trailing tilde, no space before it |
| (( )) | Unintelligible speech - double parentheses, no inner spaces |
| ((word)) | Best-guess / low-confidence transcription of a barely intelligible word |
Rule: never insert a non-speech tag mid-word - place it immediately before the word in progress. Repeated instances of the same sound in one segment are represented only once.
Segmentation rules (quick reference)
- Five primary segment types: Speech, Babble, Overlap, Music, Noise. Each segment has exactly one primary type.
- Timestamps are positive floats in seconds.milliseconds (e.g. 12.345).
- Trim continuous silence/white noise ≥2s from the edges of a segment - keep segments tight around the target sound.
- Speech segments should not exceed 15 seconds - this workbench enforces it by auto-splitting on commit; use Split/Merge to fine-tune boundaries.
- Transcription is required only for Speech (and Babble/Overlap where applicable) segments.
- Single-channel files: segments with no speech or only non-speech sound on that channel should be dropped, not transcribed as [no-speech].
Convert / Resample
Shrink oversized video, extract audio, resample speech files and convert common media formats with local FFmpeg. Originals are never overwritten.
Source media
Choose any file in the active project. For a 2-3 minute 900 MB interview, start with the 720p H.264 preset.
Quick preset
Safe presets for interviews, ASR and review copies. You can refine the settings below.
Conversion settings
Converted outputs
Generated files are added to the same project as new media and remain linked to their source.
Export
Export the current session in the format your pipeline needs. Everything runs locally in the browser.
QA summary before export
Roadmap - one console, every modality
The metadata model (file-level values + annotator + typed segments) and the export pipeline are designed to extend beyond audio without a rewrite.
🖼 Photo / image labeling
Next upType.name: "IMAGE_ANNOTATION". Replaces time-based Segments with region-based Annotations.
- Bounding boxes, polygons, and segmentation masks
- Classification tags & multi-label scene/object tags
- Attribute tags (occlusion, truncation, blur, low-light)
- EXIF / capture metadata
🎬 Video labeling
PlannedType.name: "VIDEO_ANNOTATION". Segments gain a spatial component per frame range.
- Temporal action segments - same StartTime/EndTime pattern
- Object tracking IDs (reuses SpeakerId pattern as TrackId)
- Shot/scene boundary tags, camera motion tags
💬 Chatbot prompt / conversation labeling
PlannedType.name: "CONVERSATION_ANNOTATION". Segments become conversation turns.
- Turn-level role tags, intent & topic tags
- Response quality rubric scoring, A/B preference ranking
- Safety/policy category tags, red-team flags
🧩 Shared platform layer
Foundation - live todayWhat every future tab shares with the audio workbench built today:
- Guidelines/AddendumName versioning per file
- AnnotatorId + QA warning pipeline
- Consistent native-JSON envelope
- Same multi-format export layer
Projects
Each client job is a project. Upload media, run transcription on the server, then edit in Label Audio.
New project
All projects
Sign in to load projects.
Add media - select a project below first
Jobs
Server-side transcription. Auto-refreshes every 3 seconds while this tab is open.
Sign in to load jobs.
Voices & Consent
Every cloned voice must be linked to a signed release. Reference-audio upload is blocked server-side without recorded consent.
Register a voice
Registry
Sign in to load the registry.
Users
Admin only. Roles: admin, reviewer, annotator, client.
Add user
All users and credits
Service credit rates Internal Dedomena Labs units, configurable by admin
Loading rates...
