SonicVox® stands on the shoulders of the open-source community. Every library, model, and dataset below is licensed for commercial use, and where attribution is required we list the citation here.
The 15 utility tools at /app/tools are powered by FFmpeg and a handful of audio/ML libraries. Every component here is commercial-use safe. The tools service (sonicvox-utils-api) and the voice-enhancement service both run an FFmpeg we compile ourselves under strict LGPL flags; several other services do not yet — see the note on FFmpeg below.
The workhorse: every encode, decode, filter, and mux behind our tools.
The Studio Tools service and the voice-enhancement service compile FFmpeg 7.1 from source with --disable-gpl --disable-version3 --disable-nonfree — the strictest LGPL configuration, linking no GPL-only codecs (libx264, librubberband) and no patent-encumbered ones (libfdk-aac). Several other services — including video dubbing, stem splitting, transcription and most of the speech engines — currently run their distribution's stock FFmpeg, which is built with --enable-gpl. Those images are used only to run the hosted service and are not distributed; migrating them to the same LGPL-only build is in progress. Ask legal@sonicvox.ai if you need the per-service build flags for a compliance review.
Powers the "Multilingual" TTS model (30 languages, 48 kHz) and is the primary voice-design engine that designs the voice catalog from natural-language prompts.
Powers Voice Enhancement (MossFormer2_SE_48K + MossFormer2_SR_48K bandwidth-extension pass) and Speaker Separation (MossFormer2_SS_16K). Upstream model cards state these weights were trained on "large-scale datasets including open-sourced and private data" — Alibaba has not published the corpus. Apache 2.0 covers our use of the weights without restriction, but per-dataset provenance cannot be independently audited. This is a sales-friction item for enterprise contracts, not a license defect.
Forked from upstream v3.0.0 (2026-02). Our modifications, per Apache 2.0 §4(b): a custom FastAPI server (api.py) and API-only container; 3-tier direct-media detection so S3 presigned URLs bypass yt-dlp; a Rich-Live compatibility shim for headless operation; SONICVOX_* environment overrides across the pipeline; remote WhisperX/faster-whisper ASR backends; six subtitle visual styles; a watermark overlay; minimum-subtitle-duration enforcement; and Indic/low-resource translation routing. Email legal@sonicvox.ai for the full per-file NOTICE (53 files modified, 39 SonicVox-only).
Neural machine translation across 400+ languages; serves the mt-translate service and is the translation engine for Indic and other targets where the LLM translator is weak.
Apache 2.0
Attribution
Speech-to-Text & diarization engines
Transcription, subtitle, and speaker-labeling backends. Several NVIDIA models are served by default under CC-BY-4.0, which mandates the attribution lines shown below.
Who-spoke-when segmentation for speech-to-text speaker labels, dubbing, and the multi-speaker speech-separation router. Runs with SonicVox's own MIT-licensed speaker-embedding model.
MIT does not require public attribution; listed for completeness.
Speech recognition for the Asian-4 route (zh / yue / ja / ko), with a Whisper fallback.
FunASR toolkit is MIT; the FunAudioLLM/SenseVoiceSmall weights ship under a custom “model-license” permitting attributed commercial use. See THIRD_PARTY_LICENSES.md for the full audit.
Generates the sound effects at /app/sound-effects (44.1 kHz stereo, up to 47 s). Trained on commercially-clean audio — Freesound + FMA under CC0 / CC-BY / CC-Sampling+, with no AudioSet content.
Per the Stability AI Community License we display the required attribution: “Powered by Stability AI.” The license grants free commercial use while aggregate annual revenue (counted across affiliates, and whether or not it derives from this model) stays below US$1M. At or above that threshold the license terminates and continued use requires a separate license from Stability AI, granted at their discretion.
Shipped to your browser — the subtitle preview renderer
Opening the caption design studio downloads ~4 MB of WebAssembly from sonicvox.ai/jassub/. That is a distribution of copyleft-licensed software, so the notices are served alongside the files themselves: NOTICE.txt (all components, licences, source links and our written source offer) and LICENSE.jassub.txt.
Renders the live subtitle preview in the caption design studio using the same libass engine our server-side burn-in uses, which is what makes the preview pixel-accurate. Loaded on demand from /jassub/ when the studio mounts — never part of the main bundle.
Upstream declares the composite build as “LGPL-2.1-or-later AND (FTL OR GPL-2.0-or-later) AND MIT AND MIT-Modern-Variant AND ISC AND NTP AND Zlib AND BSL-1.0”. The copyleft components are FriBidi (LGPL-2.1-or-later) and FreeType (used under the FreeType License, not the GPL option); libass itself is ISC and the JASSUB wrapper is MIT. We convey the artifacts unmodified, and the corresponding source is available upstream or from legal@sonicvox.ai at no more than the cost of distribution — see /jassub/NOTICE.txt and /jassub/LICENSE.jassub.txt, which are served publicly beside the modules they cover.
The application stack our product is served from — every item here is MIT or Apache. This is not the full platform inventory: our internal-only observability and async tooling (Grafana, Loki, Promtail — AGPL-3.0; Inngest server — SSPL; Redis 7.4 — RSALv2/SSPLv1) is unmodified, never exposed to customers as a service, and recorded with its disposition in THIRD_PARTY_LICENSES.md.