Credits
ScribeClear is built on open models made by researchers and companies who share their work. We run them ourselves, or through Cloudflare Workers AI, inside our own processing pipeline. Thank you to everyone below.
Naming these models doesn't mean their creators endorse or are connected with ScribeClear.
Indic-Transcribe-core
Built with Indic-Transcribe-core from Bodhan AI / AI4Bharat.
- What it does here:
- transcribes Hindi.
- Made by:
- Bodhan AI / AI4Bharat.
- Changes:
- none. We use the model as published.
NVIDIA Canary-1B-v2
- What it does here:
- it's the base model that Indic-Transcribe-core was built from.
- Made by:
- NVIDIA.
- Changes:
- we don't use or change it directly. Bodhan AI / AI4Bharat built Indic-Transcribe-core from it.
Koyal-hi-120M
- What it does here:
- places each word of a Hindi transcript at the right moment in the audio, and transcribes Hindi when Indic-Transcribe-core isn't available.
- Made by:
- Adalat AI. Its encoder starts from AI4Bharat's MahaDhwani pretrained model (MIT licence).
- Changes:
- none. We use the model as published.
OpenAI Whisper large-v3-turbo
- What it does here:
- transcribes every other language, on Cloudflare Workers AI. When that's unavailable, a backup provider (Groq) runs Whisper large-v3-turbo, or Whisper large-v3 for “transcribe straight to English”.
- Made by:
- OpenAI. Paper: Robust Speech Recognition via Large-Scale Weak Supervision (Radford, Kim, Xu, Brockman, McLeavey, Sutskever, 2022).
- Licence:
- MIT. Copyright (c) 2022 OpenAI.
NVIDIA Nemotron-3-Diarization
- What it does here:
- speaker labels: works out who spoke when, for up to 8 speakers.
- Made by:
- NVIDIA.
Silero VAD
- What it does here:
- finds where people are speaking, so silence and music are never sent for transcription, and cuts the audio at pauses so no word is split.
- Made by:
- the Silero Team.
- Licence:
- MIT. Copyright (c) 2020-present Silero Team.
Gemma 4 (26B A4B)
- What it does here:
- translates transcripts and subtitles, writes summaries and answers questions about a recording, on Cloudflare Workers AI.
- Made by:
- Google.
- Licence:
- Gemma 4 licence
- Source:
- ai.google.dev/gemma
DeepFilterNet
- What it does here:
- “restore audio”: makes a cleaned copy of a noisy recording for listening and download. Transcription always uses your original audio.
- Made by:
- Hendrik Schröter and contributors.
- Licence:
- MIT or Apache License 2.0, at the user's choice.
Open-source software
ScribeClear also relies on open-source software, including:
- FFmpeg (ffmpeg.org), to convert audio and video;
- yt-dlp (github.com/yt-dlp/yt-dlp, Unlicense), to import audio from links;
- faster-whisper (github.com/SYSTRAN/faster-whisper, MIT, Copyright (c) 2023 SYSTRAN), whose speech-detection code we adapted;
- NVIDIA NeMo (github.com/NVIDIA/NeMo, Apache License 2.0), to run speaker labels and the Hindi models;
- PyTorch (pytorch.org, BSD-style licence).