Skip to main content

Frequently Asked Questions

Common questions and answers about OpenTranscribe.

General​

What is OpenTranscribe?​

OpenTranscribe is an open-source, self-hosted AI-powered transcription and media analysis platform. It uses state-of-the-art AI models (Whisper via faster-whisper, a PyAnnote-based speaker engine, LLMs), or cloud providers you choose, to transcribe audio/video files, identify speakers, generate summaries, and enable powerful search across your media library.

Is OpenTranscribe free?​

Yes! OpenTranscribe is completely free and open-source under the GNU Affero General Public License v3.0 (AGPL-3.0). There are no subscription fees, usage limits, or hidden costs. You can use it for personal, commercial, or any other purpose, as long as you comply with the AGPL-3.0 license terms.

What makes OpenTranscribe different from other transcription services?​

  • Self-hosted - Your data stays on your infrastructure unless you choose a cloud provider
  • Open source - Full transparency and customizability
  • Privacy-first - No cloud services required; cloud transcription, speaker detection and LLM providers are all optional
  • Advanced features - Speaker diarization, cross-video speaker matching, AI summarization
  • No usage limits - Transcribe unlimited content
  • GPU acceleration - Fast processing with your own hardware
  • Offline capable - Works in airgapped environments

Can I use OpenTranscribe commercially?​

Yes! The AGPL-3.0 license permits commercial use. However, if you modify OpenTranscribe and offer it as a network service (SaaS), you must make your modified source code available to your users under the same AGPL-3.0 license.

Installation & Setup​

What are the system requirements?​

Minimum:

  • 8GB RAM
  • 4 CPU cores
  • 50GB disk space
  • Docker & Docker Compose

Recommended:

  • 16GB+ RAM
  • 8+ CPU cores
  • 100GB+ SSD
  • NVIDIA GPU with 8GB+ VRAM (RTX 3070 or better)

See Hardware Requirements for detailed recommendations.

Do I need a GPU?​

No, but highly recommended for practical use when transcribing on your own server. CPU-only processing is very slow:

  • GPU (RTX 3080): 1-hour video → ~5 minutes
  • CPU (8-core): 1-hour video → ~60 minutes

Speaker detection on your own server also runs without a GPU, more slowly. On a CPU-only install nothing switches it off for you (the old ENABLE_DIARIZATION variable is gone): to skip it, choose Off under Settings → Speaker Identification → Speaker Detection.

If you have no GPU at all, the lite deployment sends transcription and speaker detection to cloud providers you configure; see Deployment variants.

Can I use Apple Silicon (M1/M2/M3)?​

Yes, with the --lite image — but not with MPS acceleration. Docker Desktop on macOS has no Metal/GPU passthrough into the Linux VM the container runs in, so every container on Apple Silicon is CPU-only regardless of the host chip's MPS capability. Lite is cloud-only for transcription and speaker detection: configure a cloud transcription provider (and pyannote.ai, or a provider that returns speakers, for speaker detection); only voiceprints are computed on the Mac's CPU. See Deployment variants for which providers return speakers. (An MPS code path exists in the diarization engine, but it is unreachable in every containerized deployment; see backend/app/transcription/CLAUDE.md.)

What GPUs are supported?​

Any NVIDIA GPU with CUDA support:

  • Minimum: GTX 1060 (6GB VRAM)
  • Recommended: RTX 3070 or better (8GB+ VRAM)
  • Best: RTX 4090, A6000, A100

AMD GPUs are not currently supported (ROCm support planned for future).

How do I get a HuggingFace token?​

  1. Create free account at huggingface.co
  2. Go to Settings → Access Tokens
  3. Click "New token", select "Read" access
  4. Accept the agreement for pyannote/speaker-diarization-community-1 — this is the only repo OpenTranscribe actually gates on (auto-approved, CC-BY-4.0)
  5. Copy token to your .env file

See HuggingFace Setup for detailed instructions.

Why do I need a HuggingFace token?​

The PyAnnote speaker diarization models are "gated" - they require accepting a user agreement before downloading. The token authenticates that you've accepted the terms. Both engines that detect speakers on your server (diar-native and the in-process PyAnnote engine) are built from those models.

Without a token: Transcription works, but your server cannot load its speaker detection models, so speaker detection on this server fails with a message about the missing model access. Speaker detection by a cloud provider (pyannote.ai, or a transcription provider that returns speakers) does not need the token.

Why did the object store change in 0.6.0? Is my data safe?​

The bundled S3 object store (the minio service) now runs Silo, a maintained fork of MinIO. MinIO removed its public images, its community repository is archived, and its last community release line carries published HIGH-severity advisories with no community fix. Silo keeps MinIO's settings, routes and on-disk format, so your data volume, .env and bucket names are unchanged and nothing is copied. Back up the database and snapshot your media volume before upgrading, because a downgrade back to MinIO is not guaranteed safe. See Object storage and Upgrading to 0.6.0.

Can I keep using MinIO?​

Not the bundled image: it cannot be pulled any more, and the last community release has unpatched advisories. You can point OpenTranscribe at your own S3-compatible store with STORAGE_BACKEND=s3 instead. See Object storage.

Can I run OpenTranscribe offline?​

Yes! After initial setup and model downloads, OpenTranscribe works completely offline. Use the offline Docker Compose configuration:

docker compose -f docker-compose.yml -f docker-compose.offline.yml up -d

See Offline Installation for details.

Features & Usage​

What file formats are supported?​

Audio: MP3, WAV, FLAC, M4A, OGG, AAC Video: MP4, MOV, AVI, MKV, WEBM, FLV

Maximum file size: 15 GB by default (MAX_UPLOAD_BYTES, configurable by the administrator)

What languages are supported for transcription?​

100+ languages are supported for transcription via Whisper. See the Whisper documentation for the full list.

v0.2.0 Language Features:

  • Source Language Selection: Auto-detect or manually specify the audio language
  • Translation Toggle: Choose to keep original language or translate to English
  • Word-Level Timestamps: the built-in model produces word timestamps for every language, directly during transcription (there is no separate alignment step). Cloud providers return word or segment timing depending on the provider

Configure language settings in: Settings → Transcription → Language

Note: English transcription quality is best. Other languages work but accuracy varies.

What languages is the UI available in?​

12 languages (every release ships all 12 in lockstep):

  • English (default)
  • Arabic (\u0627\u0644\u0639\u0631\u0628\u064a\u0629) \u2014 right-to-left
  • German (Deutsch)
  • Spanish (Espa\u00f1ol)
  • French (Fran\u00e7ais)
  • Italian (Italiano)
  • Japanese (\u65e5\u672c\u8a9e)
  • Korean (\ud55c\uad6d\uc5b4)
  • Dutch (Nederlands)
  • Portuguese (Portugu\u00eas)
  • Russian (\u0420\u0443\u0441\u0441\u043a\u0438\u0439)
  • Chinese (\u4e2d\u6587)

Change the UI language in: Settings → Profile & Security (Language)

Want to contribute a translation? Submit a PR with a new locale file!

How accurate is the transcription?​

Accuracy depends on:

  • Audio quality - Clear audio with minimal background noise
  • Model size - large-v2 is most accurate
  • Language - English has best accuracy
  • Speaker accent - Native accents perform better

Typical accuracy: 85-95% word accuracy with good audio quality.

How does speaker diarization work?​

You choose who detects speakers, separately from who transcribes: your own server (the diar-native engine, or the in-process PyAnnote engine, chosen by an administrator), the cloud transcription provider when it returns speakers, or the pyannote.ai cloud service. On your server the engine:

  1. Detects voice activity - Finds when people are speaking
  2. Extracts voice features - Creates voice fingerprints
  3. Clusters speakers - Groups similar voices
  4. Assigns labels - Tags each segment with SPEAKER_00, SPEAKER_01, etc.

The speaker turns are then matched to the transcript by word timestamps (or by segment, when the model only returns segment timing). Voiceprints are always computed on your server, whoever detected the speakers. You can then manually edit speaker names, and OpenTranscribe will remember those voices across future transcriptions. See Speaker Diarization.

Can OpenTranscribe identify speakers by name automatically?​

Not automatically, but it can:

  1. Suggest speaker identities using LLM analysis of conversation content
  2. Match voices across videos using voice fingerprints
  3. Remember speakers once you've identified them in one video

See Speaker Management for details.

Why are some speaker labels corrected automatically?​

OpenTranscribe applies diarization boundary correction to fix speaker mislabeling at turn boundaries. Word-boundary smoothing (default on) collapses short "wrong-speaker islands," and an optional acoustic re-check re-embeds short disputed words and reassigns them by voiceprint. This relabels existing words only -- it never invents speech. See Boundary Correction.

Can OpenTranscribe redact PII, profanity, or toxic content?​

Yes. Content redaction detects PII, profanity, and toxic/offensive words and masks them with [CATEGORY] placeholders at every display and export surface. Masking is a read-time transform -- the full original transcript is always kept in the database. It is a per-user feature, off by default (opt-in) (Settings → Privacy & Redaction → Content Redaction), with an admin enforcement floor (Redaction Policy) that can force categories on and mandate censored exports for all users. See Content Redaction.

How many speakers can it detect?​

The default diar-native engine counts speakers automatically. You can also give a speaker range (1 to 50) in Settings → Speaker Identification → Speaker Detection, or per file when you upload or reprocess; the system default range is MIN_SPEAKERS/MAX_SPEAKERS in .env (1-20).

The range is a hint: pyannote.ai and the in-process PyAnnote engine apply it, cloud transcription providers apply what they support, and the default diar-native engine ignores it. Accuracy decreases with very large groups (30+ speakers).

What LLM providers are supported?​

OpenTranscribe supports multiple LLM providers for AI summarization:

Cloud Providers:

  • OpenAI (GPT-4o, GPT-4o-mini, GPT-4-turbo)
  • Anthropic Claude (Claude 3.5 Sonnet, Claude 3 Opus)
  • OpenRouter (access to 100+ models)

Self-Hosted:

  • vLLM (fast self-hosted inference)
  • Ollama (local models)
  • Custom OpenAI-compatible APIs

See LLM Integration for setup instructions.

Do I need an LLM for transcription?​

No! LLMs are optional and only used for:

  • AI-powered summarization
  • Topic extraction
  • Enhanced speaker name suggestions

Transcription and speaker diarization work without any LLM.

Can I generate AI summaries in languages other than English?​

Yes! AI summaries can be generated in 12 languages (LLM_OUTPUT_LANGUAGES, backend/app/core/constants.py):

  • English, Spanish, French, German
  • Italian, Portuguese, Dutch, Russian
  • Chinese, Japanese, Korean, Arabic

Configure in: Settings → Transcription → Language (Language for AI features)

The LLM will generate the summary in your chosen language regardless of the original audio language.

Can I use OpenTranscribe without any cloud services?​

Yes! You can run completely offline and local:

  • Transcription: The built-in Whisper models (choose the built-in model under Settings → Transcription)
  • Speakers: Detected on your server (choose Automatic with the built-in model, or Always on this server's GPU)
  • Voiceprints: Always computed on your server
  • LLM (optional): Local vLLM or Ollama

This needs the full deployment: a lite deployment cannot transcribe or detect speakers on the server.

No data leaves your infrastructure.

How long does processing take?​

Depends on hardware and content length:

Content LengthGPU (RTX 3080)CPU (8-core)
5 minutes~30 seconds~5 minutes
30 minutes~3 minutes~30 minutes
1 hour~5 minutes~60 minutes
3 hours~15 minutes~3 hours

Processing speed: the measured throughput on an RTX A6000 is roughly 35-39x realtime for the full pipeline, including speaker diarization (see docs/BENCHMARK_RESULTS.md in the repository). The table above is a rough estimate for a mid-range GPU; real times depend on the model, the diarization settings, the content and your hardware.

Can I process multiple files at once?​

Yes! OpenTranscribe uses Celery workers to process multiple files in parallel:

  • Default: 1-2 files at once (depending on GPU memory)
  • Multi-GPU: 4+ files at once with Multi-GPU Scaling

How much disk space do I need?​

For OpenTranscribe:

  • Docker images: ~5GB
  • AI models: several GB (depends on WHISPER_MODEL; see HuggingFace setup)
  • Database: ~100MB (grows with transcriptions)

For media files:

  • Plan ~10% of original file size for processed data
  • Example: 100 hours of audio (10GB) → ~11GB total storage needed

Can I download videos from YouTube and other platforms?​

Yes! OpenTranscribe supports the many sites that yt-dlp supports (well over a thousand extractors):

  • Best supported: YouTube (including playlists), Dailymotion, Twitter/X
  • Limited support: Vimeo, Instagram, Facebook, TikTok (may require authentication)
  • Automatically downloads and transcribes
  • User-friendly error messages for authentication-required videos

How do I view system statistics?​

System statistics (CPU, memory, disk, GPU usage) are visible to all authenticated users in the navbar - look for the system stats icon. This shows real-time resource usage of your OpenTranscribe server.

How does pagination work for large transcripts?​

For transcripts with thousands of segments (3+ hour recordings), OpenTranscribe automatically paginates the display to prevent browser slowdown. Segments load progressively as you scroll through the transcript.

Can I share transcriptions with specific groups of users?​

Yes! OpenTranscribe supports User Groups and Collection Sharing. Admins can create groups of users and then share entire collections with those groups. This makes it easy to organize team access — for example, share a "Legal Team" collection with a "Legal" user group so all members can access the same transcriptions.

Can I use cloud speech-to-text instead of local Whisper?​

Yes! OpenTranscribe supports multi-provider cloud ASR in addition to the built-in Whisper models. Speaker detection is a separate choice: most cloud providers also return speakers, and with Automatic their labels are used. For a model that returns text only, your server detects the speakers (full deployment), or you choose pyannote.ai. A model that returns no timestamps at all cannot be combined with speaker detection.

A lite deployment needs no GPU and is cloud-only: transcription and speaker detection both use cloud providers you configure, while voiceprints are still computed on your server.

When you use a cloud provider, your audio is sent to that vendor and processed under its terms; the same applies to pyannote.ai. See Deployment variants for which providers return speakers and which combinations are valid.

Can I reprocess just the speaker diarization without re-transcribing?​

Yes! The selective reprocessing feature lets you pick specific stages to re-run. Instead of reprocessing an entire file from scratch, you can choose to redo only speaker diarization (Re-diarize Only (v4)), only summarization, or any other individual stage — saving significant time when you just need to tweak one part of the pipeline.

Re-diarizing uses your current speaker-detection choice: pyannote.ai if you chose it, otherwise your server's engine (a cloud provider's own speakers cannot be fetched again without transcribing again). On a lite deployment a re-run that would need your server's engine is refused. The reprocess dialog currently offers this stage only while your active transcription provider is the built-in model.

How does cross-video speaker matching work?​

OpenTranscribe uses GPU-accelerated speaker pre-clustering to group similar voices across all your transcriptions. Voice embeddings (voiceprints) are computed on your server after speakers are detected, whoever detected them, and then clustered so the system can suggest matches when the same speaker appears in different recordings. This powers the cross-video speaker identification workflow.

Can OpenTranscribe automatically tag my transcriptions?​

Yes! The AI auto-labeling feature uses your configured LLM to analyze transcription content and suggest relevant topics. These suggestions are automatically applied as tags and can be organized into collections, making it easy to categorize and find related content without manual effort.

Can files be automatically deleted after a certain period?​

Yes! Admins can configure file retention policies to automatically clean up old files after a specified number of days. This is useful for compliance requirements or simply managing disk space. Retention policies are configured in the admin settings and apply system-wide.

How can I improve AI summary quality for my organization?​

Configure Organization Context in Settings → Organization Context. This lets you provide background information about your organization — such as its mission, common terminology, and typical meeting types. The LLM uses this context when generating summaries, resulting in more relevant and accurate output tailored to your specific domain.

Can I switch between grid and list view?​

Yes! The gallery view supports both grid and list layouts. You can toggle between them using the view switcher in the toolbar. Your preference is remembered between sessions. The gallery also includes pagination for efficiently browsing large libraries.

Does OpenTranscribe detect speaker gender?​

Yes! OpenTranscribe includes automatic gender classification for detected speakers. Gender predictions are made from each speaker's audio by a voice-attribute classifier (a Wav2Vec2-based model, or the diar-native engine in the same pass when it detected the speakers) and displayed alongside speaker information. This helps validate speaker clusters and provides additional metadata for organizing and identifying speakers.

Can I customize export format?​

Yes! The TXT export feature includes configurable options that let you control what's included in the exported file — such as timestamps, speaker labels, and formatting preferences. Your export settings are saved as persistent preferences so you don't have to reconfigure them each time.

Performance & Optimization​

How can I speed up processing?​

  1. Use a GPU - Fastest option (required for practical use)
  2. Use larger batch size - If you have GPU memory (edit BATCH_SIZE in .env)
  3. Use smaller model - medium or base (faster but less accurate)
  4. Multi-GPU scaling - Process 4+ files in parallel
  5. SSD storage - Faster disk I/O

My GPU is running out of memory. What can I do?​

  1. Reduce batch size: BATCH_SIZE=8 (or lower)
  2. Use smaller model: WHISPER_MODEL=medium or base
  3. Use quantization: COMPUTE_TYPE=int8
  4. Close other GPU applications
  5. Upgrade GPU (if feasible)

Can I use multiple GPUs?​

Yes! OpenTranscribe supports multi-GPU scaling for high-throughput processing:

# Configure in .env
GPU_SCALE_ENABLED=true
GPU_SCALE_DEVICE_ID=2 # Which GPU to use
GPU_SCALE_WORKERS=4 # Number of parallel workers

# Start with GPU scaling
./opentr.sh start dev --gpu-scale

See Multi-GPU Scaling for details.

Troubleshooting​

OpenTranscribe won't start​

# Check Docker is running
docker ps

# Check logs
docker compose logs

# Common issues:
# - Port conflicts (5173, 8080, etc. already in use)
# - Insufficient memory
# - Docker Compose not installed

See Troubleshooting Guide.

"Permission denied" error for model cache​

# Fix permissions
./scripts/fix-model-permissions.sh

# Or manually
sudo chown -R 1000:1000 ./models

This happens because Docker creates directories as root, but containers run as non-root user (UID 1000) for security.

Transcription fails with "CUDA out of memory"​

Reduce GPU memory usage:

# Edit .env
BATCH_SIZE=8 # Reduce from 16
COMPUTE_TYPE=int8 # Use 8-bit quantization

# Or use smaller model
WHISPER_MODEL=medium

Speaker diarization not working​

  1. Check your choice under Settings → Speaker Identification → Speaker Detection (it may be Off), and check whether the file shows a refusal message (for example a model with no timestamps, or pyannote.ai chosen without a key)
  2. For detection on your server: check the HuggingFace token is set in .env
  3. Accept the model agreement for pyannote/speaker-diarization-community-1
  4. Check logs for download errors: docker compose logs celery-worker
  5. Re-download models if corrupted

On a lite deployment there is no speaker detection on the server; choose a cloud model that returns speakers or pyannote.ai.

Poor transcription quality​

  • Use better audio - Clear, well-recorded audio
  • Use larger model - large-v2 is most accurate
  • Reduce background noise - Use noise cancellation
  • Check language setting - Ensure correct language selected

YouTube downloads failing​

  • Check yt-dlp version - May need updating
  • Check video availability - Some videos are region-locked
  • Check age restrictions - Age-restricted videos may not work
  • Try direct file upload - Download manually first

Security & Privacy​

Is my data secure?​

Yes. By default all processing happens on your server:

  • Media files stored in MinIO (local S3-compatible storage)
  • Transcripts stored in PostgreSQL (local database)
  • AI models run locally with the built-in transcription model and speaker detection on your server
  • Voiceprints are always computed on your server
  • Optional LLM calls can use self-hosted providers

If you choose a cloud transcription provider, pyannote.ai, or a cloud LLM, the audio (or, for an LLM, transcript text) is sent to that vendor.

Are transcripts encrypted?​

It depends on the store:

  • Media files (MinIO): encrypted at rest by default — the installer enables AES-256-GCM server-side encryption automatically.
  • Credentials and secrets (API keys, TOTP secrets): encrypted at the application layer with AES-256-GCM before they reach the database.
  • Transcript text (PostgreSQL and OpenSearch): not encrypted at the application layer — search and AI features require the backend to read it. PostgreSQL has no built-in at-rest encryption, so use full-disk encryption (LUKS/dm-crypt, FileVault, BitLocker) or place the data volumes on an encrypted filesystem.
  • Backups: use ./opentranscribe.sh backup --encrypt for GPG AES-256 encrypted dumps.

See Security Hardening for the complete picture.

Who can access my transcriptions?​

Only users you create. OpenTranscribe has role-based access control:

  • Admin users - Full access to all data
  • Regular users - Access only to their own data

Can I use OpenTranscribe for sensitive content?​

Yes! OpenTranscribe is designed for privacy-sensitive use cases:

  • Legal depositions
  • Medical consultations
  • Business strategy meetings
  • Personal recordings

With the built-in transcription model, speaker detection on your server and a local (or no) LLM, nothing is sent to cloud services. Cloud transcription, pyannote.ai speaker detection and cloud LLMs are each an explicit choice that sends audio or text to that vendor.

Authentication​

What authentication methods are supported?​

OpenTranscribe supports multiple authentication methods:

  • Local - Username/password stored in PostgreSQL (default)
  • LDAP/Active Directory - Authenticate against existing AD infrastructure
  • OIDC/Keycloak - Single Sign-On via OpenID Connect
  • PKI/X.509 - Certificate-based authentication (CAC/PIV card support)

Multiple methods can be enabled simultaneously for hybrid deployments.

See Authentication Overview for details.

Does OpenTranscribe support MFA?​

Yes! OpenTranscribe supports TOTP-based multi-factor authentication:

  • Works with Google Authenticator, Authy, and other TOTP apps
  • QR code setup for easy configuration
  • Backup codes for account recovery
  • Per-user enablement

Enable MFA in your .env file:

MFA_ENABLED=true
MFA_ISSUER_NAME=OpenTranscribe

Can I integrate with Active Directory?​

Yes! OpenTranscribe supports LDAP/Active Directory authentication:

  • Users authenticate with AD credentials
  • Accounts auto-created on first login
  • Admin roles configurable via LDAP_ADMIN_USERS
  • Hybrid mode supports both local and AD users

See the LDAP Setup Guide for configuration.

What password policies are available?​

OpenTranscribe includes FedRAMP-compliant password policies:

  • Minimum length (default: 12 characters)
  • Complexity requirements (uppercase, lowercase, digits, special characters)
  • Password history (prevent reuse of last 24 passwords)
  • Password expiration (default: 60 days)
  • Common pattern detection (blocks weak passwords)

Configure in .env:

PASSWORD_POLICY_ENABLED=true
PASSWORD_MIN_LENGTH=12
PASSWORD_HISTORY_COUNT=24
PASSWORD_MAX_AGE_DAYS=60

Is there account lockout protection?​

Yes! OpenTranscribe implements NIST AC-7 compliant account lockout:

  • Lock account after failed attempts (default: 5)
  • Progressive lockout durations (15 min, 30 min, 60 min, 24 hours)
  • Admin unlock capability
  • Automatic unlock after lockout expires

Configure in .env:

ACCOUNT_LOCKOUT_ENABLED=true
ACCOUNT_LOCKOUT_THRESHOLD=5
ACCOUNT_LOCKOUT_DURATION_MINUTES=15
ACCOUNT_LOCKOUT_PROGRESSIVE=true

Does OpenTranscribe support SSO?​

Yes — via any conforming OpenID Connect provider, discovered from its .well-known/openid-configuration: Keycloak, Authentik, Authelia, Okta, Entra ID, Auth0, Zitadel, and so on.

  • Authorization-code flow with PKCE, ID-token-only validation
  • Group and role synchronization through a configurable roles claim
  • Group mapping from directory groups to in-app groups and roles
  • RP-initiated (federated) logout where the provider publishes an end-session endpoint

See the OIDC Setup Guide for configuration. Existing KEYCLOAK_* environment variables keep working permanently.

Can I use CAC/PIV cards for authentication?​

Yes! OpenTranscribe supports PKI/X.509 certificate authentication:

  • Client certificate authentication via mutual TLS
  • CAC/PIV smart card support
  • No passwords required
  • Admin designation via certificate DN

See the PKI Setup Guide for configuration.

How is audit logging handled?​

OpenTranscribe includes FedRAMP-compliant audit logging:

  • Structured JSON or CEF format
  • All authentication events logged
  • Optional OpenSearch integration for analysis
  • Tracks logins, logouts, failed attempts, MFA events, admin actions

Configure in .env:

AUDIT_LOG_ENABLED=true
AUDIT_LOG_FORMAT=json
AUDIT_LOG_TO_OPENSEARCH=false

Development & Contribution​

How can I contribute?​

We welcome contributions! See Contributing Guide for:

  • Code contributions
  • Documentation improvements
  • Bug reports
  • Feature requests
  • Translations

Can I customize OpenTranscribe?​

Yes! OpenTranscribe is open source (AGPL-3.0 License):

  • Modify the code
  • Add custom features
  • Integrate with other services
  • White-label for your organization

See Developer Guide to get started.

Important: If you modify OpenTranscribe and offer it as a network service (SaaS), you must make your modified source code available to your users under the AGPL-3.0 license.

How is OpenTranscribe built?​

Frontend: Svelte + TypeScript + Vite Backend: Python + FastAPI + SQLAlchemy AI: faster-whisper + diar-native / PyAnnote + LangChain Infrastructure: Docker + PostgreSQL + Redis + MinIO + OpenSearch

See Architecture for details.

What AI models does it use?​

  • Whisper (faster-whisper) - Speech recognition, with word timestamps produced during transcription (no separate alignment model)
  • PyAnnote community-1 - Speaker diarization, run by the diar-native engine or in-process PyAnnote
  • WeSpeaker - Voiceprints for cross-file speaker matching
  • Wav2Vec2 gender/age classifier - Optional speaker attribute detection
  • Sentence Transformers - Semantic search embeddings
  • LLMs (optional) - Summarization and analysis

What is the license?​

OpenTranscribe is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0) - a strong copyleft open-source license. You can:

  • Use commercially
  • Modify the code
  • Distribute
  • Private use

Key requirement: If you modify OpenTranscribe and offer it as a network service (SaaS), you must provide your users access to the modified source code under the same AGPL-3.0 license. This ensures the open source community benefits from improvements.

Can I sell OpenTranscribe?​

Yes! The AGPL-3.0 license permits commercial use, including:

  • Offering it as a paid service
  • Selling access to your installation
  • Using it for commercial transcription work

Important: If you modify the code and run it as a network service, you must make your modifications available under AGPL-3.0.

Do I need to credit OpenTranscribe?​

Credit is appreciated but not required. The AGPL-3.0 license requires that you:

  1. Include the original license notice in copies of the software
  2. Make source code available if you offer it as a network service
  3. Preserve copyright notices

What if I don't want to share my modifications?​

If you modify OpenTranscribe for internal use only (not as a network service), you don't need to share your changes. The AGPL-3.0 only requires source disclosure when you offer the software as a service to others over a network.

What about the AI models' licenses?​

  • WhisperX: Apache 2.0 License
  • PyAnnote: MIT License
  • Wav2Vec2 gender/age classifier: Apache 2.0 License

All are permissive licenses compatible with commercial use.

Still Have Questions?​