AI Policy & Transparency
AI ipi tinoshandisa, uko inofamba, uye sei tichitevedzera zvinodiwa zvekuzivisa (EU AI Act Chikamu 50, chinotanga kushanda 2026-08-02).
TL;DR
- Zvese transcripts ari AI-yakagadzirwa. Every output inotakura machine-inoverenga kuratidzwa.
- Isu hatisi kutumira yako audio kana chinyorwa ku OpenAI, Anthropic, Google, kana chero mumwe LLM API.Izvi zvinoitika nekuti isu tinoda kuti iwe uzive kuti ndeapi masevhisi edu anoshanda.
- Isu hatina kudzidzisa chero model pa transcripts yako. Every model isu kutamba ndiro off-the-shelf open-source model. Kugadzirisa iwe zvakajeka kuita zvinounganidzwa chete kana iwe kuvhota mu.
- Synthetic voice clones (TTS) zvakajeka zvinonzi AI-yakagadzirwa muzita refaira, metadata, uye papeji.
Models isu kushandisa
| Model | Yakashandiswa | License | Inofamba |
|---|---|---|---|
| Whisper large-v3-turbo (faster-whisper) | Kunyora (default) | MIT | Yedu GPU |
| Whisper large-v3 (faster-whisper) — branded “STT.ai Enhanced” | Transcription (zvirongwa zvakabhadharwa) | MIT | Yedu GPU |
| Vosk | Real-time word streaming | Apache 2.0 | Yedu GPU |
| SpeechBrain ECAPA-TDNN | Mutauro wemutauro | Apache 2.0 | Yedu GPU |
| MadLAD-400 3B | Kushandurwa (450+ languages) | Apache 2.0 | Yedu GPU |
| Qwen2.5-1.5B (llama.cpp) | Summary, analysis, content generation, RAG chat | Apache 2.0 | Yedu GPU |
| Chatterbox (Resemble AI) | Voice cloning / text-to-speech | MIT | Yedu GPU |
| all-MiniLM-L6-v2 | Embeddings for RAG search | Apache 2.0 | Yedu GPU |
“STT.ai Enhanced” ndechimwe chezvigadzirwa zvedu zvinonzi Whisper Large-v3, izvo zvinoshanda paGPU yedu. Hachisi chigadzirwa chedu chekuzvigadzirisa, chine hunyanzvi, kana chine hunyanzvi, asi chikuru uye chakanyanya kunyatsoshanda. Chiri kunyatsoshanda kupfuura default Whisper Large-v3-turbo, uye chinosvika 1.5x chakanyanya kuomarara pabasa rakanyorwa, rakaongororwa pahardware yedu.
Isu hatisi kukumbira OpenAI, Anthropic, Google Cloud, kana chero wechitatu-bato LLM API ye chero chinhu. Every model pamusoro rinofamba pa hardware isu tiri pamutemo uye kudzora. The chete kunze AI sevhisi isu kushandisa translateapi.ai (aka Muddy Holdings) kushandura UI zvinyorwa — izvi kamwe kumbobvira kusvikira yako translitterated zvinhu.
Maitiro edu ekuzivisa AI generation
- Transcript HTML mapeji: Kusanganisira <meta name=\
- Kutumira kunze kwemashoko (TXT, SRT, VTT, JSON, CSV, DOCX, PDF): kusanganisira 'AI-yakagadzirwa transcript' header line pakutanga kweese faira.
- Sintetic voice / TTS output: WAV mafaera anosanganisira'synthetic-voice' tag mumetadata uye chinoratidzwa chinoratidza kuti faira iri kurodha pasi. Kutaura nezvekuzvidzivirira kunogona kutaurwa munzira.
- API mazano: Kusanganisira _ai_generated: true field muJSON response iyo ine transcribed content.
Kudzidzisa data
- Base models (Whisper, MadLAD, Qwen, etc.) vanouya vadzidziswa kubva kune vaparidzi vavo. Tinoshandisawo mamodheru aya sezvavanouya.
- Transcripts yako haasi kushandiswa kudzidzisa base mamodheru.
- Kana iwe ukagadzirisa transcript segment (pencil icon) kana ukaratidza kuti haina kunaka (flag icon), AND iwe wakasarudza ku /privacy-settings/ ("Model Training" toggle, default off), the (incorrect text → corrected text) pair may be collected to train a future text-correction model. Only the text pair — never the surrounding audio — is used by this path. No such model exists today: the pairs are only stored, and nothing we currently serve has been trained on them.
- Zvichisiyana, Contribute corrections plus audio to Voice Lab toggle (akaiswawo pa /privacy-settings/, akaiswawo pasi se default) inokutendera kuti uite zvikonzero zvekushandura ma segments ako, pamwe nemashoko akagadziridzwa, kune yedu Voice Lab dataset pasi peCC-BY-SA-4.0. Ma toggles maviri aya anoshanda zvakazvimirira — unogona kubvumira chero chinhu, zvese, kana chero chinhu.
- Audio yauno upload inobvisa mukati me 24 hrs nenguva yenguva yekugadzirisa basa, neimwe chete yezvikonzero zviviri zvinoitika chete kana iwe uchizvisarudza. Kana iwe wakasarudza "Contribute corrections + audio to Voice Lab" pa /privacy-settings/, audio inochengetwa kwemazuva 90 kuitira kuti Voice Lab ikwanise kubvisa segment yakagadzirirwa; kana yatochengetwa (kana mushure memazuva 90, chero ipi yaive yapfuura), audio yemutauro unotanga inobvisa. Kana iwe wakapinda uye ukaisa audio yako yekurekodha mu link yekugovana, iyo audio inochengetwa kwemazuva 30 kana iwe ukawana account yemahara; 30 mazuva, 90 mazuva kana pasina nguva yemahara neForever pa plan yekubhadhara) uye inobvisa mukati me 24 hrs mushure meku expire link. Audio yakagovana haichashandiswa kudzidzisa. Transcript yakanyorwa inotevera kuchengetedza pfungwa ye plan yako (anonymous 24h, free 30 days, Pro+ permanent).
Kunyatsonzwisisa uye mashandisiro
AI transcription hachisi chakakwana. Word matambudziko mitengo zvinosiyana ne speaker accent, audio quality, rurimi, uye domain mashoko. For critical use (legal, medical, akachengeteka indasitiri) tarisa pamusoro pekutanga audio. Our public WER benchmarks per model are at /models/.
Mabvunzo ekuita
Kuti uwane ruzivo nezve EU AI Act, GDPR, kana zvimwe zvikumbiro zvekutevedzera: hello@stt.ai kana kushandisa Contact form fomu.
Last updated: 2026-08-05.