Sharciga AI & Shaandhaynta
Waa maxay AI aan isticmaalno, halka ay socoto, iyo sida aan u raacno shuruudaha shaacinta (EU AI Act Qaybta 50, 2026-08-02).
TL;DR
- Dhammaan transcripts waa AI-soo saaro. Wax soo saarka kasta oo qaado shaacin mashiinka akhris.
- AI oo dhan waxay ku socdaan GPU-yadayaga. Waxaanan u diri karin codkaaga ama qoraalka OpenAI, Anthropic, Google, ama wax kasta oo saddexaad LLM API.
- Waxaan u tababaraan ma aha nooc kasta oo ku saabsan qoraaladaada. Nooca kasta oo aan la socdaa waa off-the-shelf furan-source qaabka. Kormeerka aad si cad u sameeyaan waxaa lagu ururiyaa oo keliya haddii aad doorato in.
- Synthetic codka klones (TTS) waa si cad u calaamadsan sida AI-soo saaro in filename, metadata, iyo on page.
Models aan isticmaalno
| Nooc | La isticmaalayo | Liisan | Ku socda |
|---|---|---|---|
| Whisper large-v3-turbo (faster-whisper) | Qoraal (default) | MIT | GPU-yada |
| Whisper large-v3 (faster-whisper) — branded “STT.ai Enhanced” | Qoraalka (plans la bixiyo) | MIT | GPU-yada |
| Vosk | Wareegta ereyga waqti-dhab ah | Apache 2.0 | GPU-yada |
| SpeechBrain ECAPA-TDNN | Qof-dheer | Apache 2.0 | GPU-yada |
| MadLAD-400 3B | Tarjumaadda (450+ luqadood) | Apache 2.0 | GPU-yada |
| Qwen2.5-1.5B (llama.cpp) | Faahfaahin, falanqaynta, abuurista content, RAG chat | Apache 2.0 | GPU-yada |
| Chatterbox (Resemble AI) | Dhaqdhaqaaqa codka / qoraal-in-dhaqaaq | MIT | GPU-yada |
| all-MiniLM-L6-v2 | Embeddings for RAG search | Apache 2.0 | GPU-yada |
“STT.ai Enhanced” waa magaca alaabtayada ee Whisper weyn-v3 socda on our GPU. Waa ma aha qaabka custom, a fine-tune, ama miisaanka gaarka ah — waa mid ka mid ah ugu weyn iyo ugu saxsan Whisper checkpoint. Waa ka saxsan ka dib markii default Whisper weyn-v3-turbo, iyo ku saabsan 1.5x ka sii dhakhso badan on shaqo diarized, la miisaamiyey on our hardware.
Waxaanan u yeedhay OpenAI, Anthropic, Google Cloud, ama mid kasta oo saddexaad LLM API feature kasta. Mid kasta oo ka mid ah qaababka kor ku xusan ku socda qalabka aan leenahay iyo hawlgalka. adeegga AI oo kaliya ee dibadda ah ee aan isticmaalno waa translateapi.ai (iyo Muddy Holdings) si loo turjumayo UI strings - tani ma taabtaan content aad ku qortay.
Sida aan u muujinno AI dhalasho
- Tusaale HTML: ku dar <meta magac=\
- Qoraalka dhoofinta (TXT, SRT, VTT, JSON, CSV, DOCX, PDF): ku jiraan 'AI-soo saaro transcript' line madaxa ee korka fayl kasta.
- Codka isku dhafan / TTS soosaarka: Faylasha WAV waxaa ka mid ah'synthetic-voice' tag in metadata iyo ogeysiis cad oo ku saabsan bogga soo dejinta. Audible diidmada waa ku saabsan qorshaha socda.
- Jawaabihii API: ku dar _ai_generated: dhab ah ee goobta jawaab kasta oo JSON oo ku jira waxyaabaha la qoray.
Macluumaadka Tababarka
- Base models (Whisper, MadLAD, Qwen, iwm) yimaadaan hore-ku tababaran ka soo daabacaadda ka mid ah. Waxaan u isticmaali sida-shipped.
- Your transcripts ma aha in la isticmaalo si ay u tababaraan qaababka base.
- Haddii aad saxdo qaybta transcript (labada icon) ama calaamad u qalad ah (labada icon), IYO aad dooratay in /privacy-settings / ("Model Training" toggle, default off), (waxyaabaha aan sax ahayn → qoraalka saxda ah) labada waa in la ururiyo si ay u tababaraan mustaqbalka qoraalka-waxyaabaha qalad ah. kaliya qoraalka labada — marnaba audio ku hareeraysan — waxaa loo isticmaalaa by jidkan. No sida noocaas ah oo qaab ah maanta jira: labada waa kaliya la kaydiyaa, iyo waxba aan hadda adeegaan waxaa tababaray on iyaga.
- Si gaar ah, Contribute corrections plus audio to Voice Lab toggle (oo sidoo kale ku yaal /privacy-settings/, sidoo kale default off) wuxuu kuu ogolaanayaa inaad ku bixiso maqalka qaybaha aad saxdo, oo la xiriira qoraalka la hagaajiyey, si aan u helno xogta Voice Lab ee CC-BY-SA-4.0. Labada toggles waa madaxbannaan — waxaad siin kartaa mid, labadaba, ama midnaba.
- Audio aad soo dejisan waxaa la tirtiraa 24 saacadood gudahood by a saacaddii shaqada nadaafadda, la laba xaddidaad oo kaliya ku habboon haddii aad doorato iyaga. Haddii aad dooratay in "Contribute corrections + audio to Voice Lab" at /privacy-settings/, audio waxaa lagu kaydiyaa ilaa 90 maalmood si maalin kasta Voice Lab ingestion ka soo bixi karaa qaybta la hagaajiyey; mar inta lagu jiro (ama ka dib 90 maalmood, mid kasta oo ka hor), audio asalka la tirtiro. Haddii aad diiwaan gashan tahay oo ay ku jiraan recordings audio in share link, audio in la haystaa ilaa link waa nool (til 30 maalmood on a free account; 30 maalmood, 90 maalmood ama aan xaddidnayn la Forever on a lacagta qorshaha) iyo waxaa la tirtiro 24 saacadood gudahood ka dib link dhamaadka. Shared audio ma loo isticmaali tababarka. qoraalka transcript raacaa sharciga haysta ee qorshahaaga (anonymous 24h, free 30 maalmood, Pro + joogto ah).
Xaq iyo khaladaad
AI qoraalka ma aha mid buuxa. Word khaladaadka rates kala duwan yihiin iyadoo la isticmaalayo afka, tayada audio, afka, iyo domain afka. Si loo isticmaalo muhiim ah (dhaqanka, caafimaadka, warshadaha la xakameeyay) xaqiijin ka dhan ah audio asalka ah. Our WER guud benchmarks per model waa at /models/.
Su'aalaha waafaqsan
Xeerka EU AI, GDPR, ama su'aalaha kale ee waafaqsanaanta: hello@stt.ai ama isticmaal xiriirka foomka.
La cusbooneysiiyay: 2026-08-05.