Underlying Models
English version controls · Updated October 19, 2026
No "advanced AI models" and no "proprietary magic". Voxtakes is a pipeline of specific, named models — mostly open-weight and self-hosted on our own GPU infrastructure. Here is every stage, every model, every vendor.
| Pipeline stage | Model | Who makes it | How it runs |
|---|---|---|---|
| Speech recognition (ASR) — transcribing your source video, and re-transcription QC scoring on every dubbed line | Whisper large-v3, running on the faster-whisper engine | OpenAI (open weights, MIT); faster-whisper by Guillaumekln | Self-hosted on our GPUs |
| Translation (MT) — line-by-line translation with your glossary enforced | Hunyuan MT 2.0 7B (primary) · MiMo v2.6 flash (fallback) | Tencent — Hunyuan MT series (open weights); Xiaomi — MiMo open-weight LLM series | Both invoked through model APIs (OpenRouter / direct MiMo endpoint); every other stage is self-hosted |
| Voice cloning & dubbing (TTS) — per-line synthesis in the selected voice | IndexTTS-2.5 · Chatterbox multilingual · MiMo voice-clone route | IndexSpeech team (Bilibili) — open weights; Resemble AI — Chatterbox (MIT); Xiaomi — MiMo audio | All self-hosted on our GPUs; engine chosen per target language |
| Vocal separation — isolating the voice bed for clean cloning references and background remix | Demucs (htdemucs) | Meta AI (MIT) | Self-hosted, one pass shared across languages |
| Speaker diarization — auto detecting who speaks each line | ESNet VoiceEncoder (Chatterbox speaker encoder) | Resemble AI (MIT) | Self-hosted, CPU, ~1s per job |
| Burned-in subtitle removal — detecting and painting out hardcoded captions | RapidOCR (PaddleOCR text detection) + STTN video inpainting; LaMa is the planned upgrade path | Paddle (Baidu, Apache-2.0); STTN & LaMa — academic open source (MIT / Apache-2.0) | Self-hosted, machine-QC verified (re-OCR of cleaned frames) |
| Quality control — every dubbed line re-listened and scored before the master is assembled | Same Whisper large-v3 re-transcription scorer + per-line glossary enforcement | OpenAI weights; scoring logic ours | Self-hosted, runs inside every job |
Where your data goes
Media, voice references and audio never touch a third-party model API: ASR, TTS, separation, diarization and subtitle removal all run on GPUs we operate (RunPod Inc., US region for Standard jobs; our own GPU infrastructure in Asia for Economy jobs). The single model call that leaves our infrastructure is translation: text lines sent through OpenRouter (OpenRouter, Inc.) to Tencent Hunyuan MT 2.0 7B, with Xiaomi MiMo v2.6 flash as the fallback route. Full processor detail lives in Section 5 of our Privacy Policy, and the deletion promise (7 days, both copies) applies regardless of route.
Model changes
When a stage changes model, this page changes in the same release, and What's new says so. You signed up for a pipeline you can read, not a black box.