Khong phai tat ca cac tren tre giong noi deu bang nhau ve do tre - va do tre la ca tiem nen.
Tren tre giong noi real-time xu ly am thanh 400ms sau khi ban noi la ve mat ky thuat la real-time theo nghia no khong yeu cau pre-recording. Nhung 400ms la do tre du lon de phá huy toan bo dong luong hoi thoai, kich hoat hieu ung dan om trong tai nghe cua ban, va lam cho moi keu goi thoai giong nhu ban dang noi qua lien ket vu tru bi hong.
Huong dan nay di sau vao toan hoc do tre tren cac tren tre giong noi live tren Windows - cach che do WASAPI Ek Quyen hoat dong, cach no so sanh voi ASIO, y nghia cua duoi 100ms / duoi 300ms / duoi 500ms nguong trong thuc te, va cach cau hinh he thong cua ban de dat cac con so thap nhat co the.
Stack Do Tre: Toi Cac Miligiay Di Dau
Do tre end-to-end trong tren tre giong noi khong phai la mot con so. Dó là tong cong cua nhieu lop, moi lop them do tre cua no:
1. Do tre dua vao driver - thoi gian doc buffer am thanh tu microphone cua ban. O 128 frame / 48kHz trong WASAPI Ek Quyen: ~2.67ms.
2. Do tre driver dua ra - thoi gian viet buffer vao thiet bi dau ra cua ban. Tinh toan tuong tu: ~2.67ms.
3. Do tre xu ly am thanh - thoi gian thuat toan tren tre giong noi cua ban bien doi am thanh. Doi voi cac hieu ung DSP: 2-10ms. Doi voi chuyen doi giong noi AI: 60-180ms tuy thuoc vao phan cung.
4. Dien tich overhead tach am thanh Windows - co the bo qua trong WASAPI Ek Quyen (duong dan thiet bi truc tiep); 20-30ms trong WASAPI Dung Chung (mixer he thong); khong ap dung voi ASIO.
5. Dien tich overhead thiet bi am thanh ao - hau het cac tren tre giong noi dinh tuyen am thanh da xu ly qua tren dua ra driver microphone ao. Thiet bi ao duoc viet tot them 5-15ms. Cai duoc viet xau co the them 40-80ms.
Cong tat ca nhung dieu do va ban nhan duoc do tre end-to-end that cua ban. Hai muc dau tien duoc sua chinh boi cai dat kich thuoc buffer cua ban. Muc 4 va 5 duoc xac dinh boi che do driver cua ban va chat luong cua viec trien khai thiet bi ao tren tre giong noi.
| Cau hinh | Do tre dua vao driver | Xu ly | Tong cong (DSP) | Tong cong (AI, GPU) |
|---|---|---|---|---|
| WASAPI Dung Chung, 1024 frame | 40-60ms | 5-15ms | 60-90ms | 120-200ms |
| WASAPI Ek Quyen, 256 frame | 10-15ms | 5-15ms | 25-40ms | 80-160ms |
| WASAPI Ek Quyen, 128 frame | 5-10ms | 5-15ms | 15-30ms | 70-150ms |
| ASIO, 64 frame | 2-5ms | 5-15ms | 10-25ms | 65-140ms |
Che Do WASAPI Ek Quyen: No Lam Gi va Tai Sao Quan Trong
Windows co hai mo hinh dua ra driver am thanh ma hau het cac tren tre giong noi co the su dung: WASAPI Dung Chung va WASAPI Ek Quyen.
WASAPI Dung Chung chay qua Windows Audio Device Graph (audiodg.exe). Am thanh cua moi ung dung duoc trung hop lai trong phan mem truoc khi dat den thiet bi cung. Tren tre them nay add laten cy - thuong la 20-30ms - va ep buoc lay mau lai neu ty le lay mau cua ban khong khop voi cai dat am thanh toan he thong (mac dinh 48kHz, 16-bit tren hau het cac he thong). Neu tren tre giong noi cua ban duoc dat thanh 44.1kHz va Windows dat thanh 48kHz, thiet bi lay mau them mot vai miligiay nua va lam giam chat luong am thanh.
WASAPI Ek Quyen bo qua mixer toan bo. Ung dung cua ban chiem quyen toan bo phan cung, cau hinh no o ty le lay mau va kich thuoc buffer lua chon cua ban, va doc/ghi truc tiep. Mixer Windows khong lien quan. Dieu nay loai bo dien tich overhead 20-30ms mixer va chi phi lay mau lai. Trao doi: khong co ung dung nao khac co the su dung thiet bi am thanh do dong thoi.
Doi voi tren tre giong noi, trao doi nay gan nhu luon luon xung dang. Ban dang dinh tuyen tat ca am thanh qua thiet bi ao tren tre giong noi cua ban - cac ung dung khac gui am thanh cua ho den cac dau ra khac.
De kiem tra xem tren tre giong noi co thuc su su dung WASAPI Ek Quyen hay khong: mo Task Manager khi tren tre giong noi dang chay, tim su dung CPU audiodg.exe. Neu no cao tren ~2%, tren tre giong noi o che do Dung Chung va tra thue mixer.
ASIO: Khi No Xung Dang va Khi Khong
ASIO (Audio Stream Input/Output) la chuan tren dua ra duoc phat trien boi Steinberg cung cap truy cap thiet bi cung truc tiep, tuong tu WASAPI Ek Quyen nhung voi dieu khien cap thap hon va laten cy dat duoc thuong thap hon.
Cac khac biet thuc te cho tren tre giong noi live:
Loi the ASIO:
- Co the duy tri buffer 64 frame (1.3ms o 48kHz) mot cach dang tin cay tren phan cung hien dai
- Dien tich overhead CPU thap hon o kich thuoc buffer tuong duong
- Do tre it bien thay hon - jitter thap hon, dieu nay quan trong cho cac mo hinh AI xu ly chunk kich thuoc co dinh
Nhuoc diem ASIO:
- Yeu cau giao dien am thanh cam ket (Focusrite Scarlett, MOTU, RME, v.v.)
- Khong co san tren am thanh build-in - Realtek onboard va Intel HD Audio khong co tren ASIO that; ASIO4ALL la shim khong cung cap toan bo loi the
- Giao dien co gia $100-$600; qua muc neu ban chi muon tren tre giong noi do tre thap
- Mot so thiet bi am thanh ao khong expose giao dien ASIO, dut doan chuoi dinh tuyen
Khuyen nghi thuc te: WASAPI Ek Quyen o 128 frame la lua chon phu hop cho hau het cac nguoi dung tren tre giong noi. Su khac biet do tre giua ASIO o 64 frame va WASAPI Ek Quyen o 128 frame la khoang 1-3ms - khong phat hien duoc trong bat cu kich ban hoi thoai the gioi thuc nao. Dau tu vao ASIO neu ban cung dang lam san xuat am nhac va can no cho cong viec DAW; khong mua giao dien am thanh chi de thay doi giong noi.
Ba Tang Do Tre va Chu Y Cam Thay
Duoi 100ms: Trong Suot
O duoi 100ms end-to-end, hau het cac nguoi dung khong the canh sac bat cu do tre nao. Hoi thoai tao ro rang. Thậm chí so sanh truc tiep giua microphone thon cua ban va dau ra duoc xu ly trong cung mot hoi thoai khong phat hien su khac biet thoi gian nhat thi.
Tang nay yeu cau:
- Che do driver WASAPI Ek Quyen hoac ASIO
- Buffer 128-256 frame
- Xu ly DSP (pitch shift, formants, EQ), HOAC chuyen doi giong noi AI voi GPU roi rac
Do do trong the gioi that cho PC choi game Windows thuong voi GPU tam nhan: WASAPI Ek Quyen + 128 frame + chuyen doi giong noi AI = 85-110ms end-to-end. Gan nhu o nguong, nhung hau het cac nguoi dung bao cao cam thay khong thay.
Duoi 300ms: Co The Su Dung Duoc
Giua 100ms va 300ms, do tre tro nen canh sac trong giam sat tai nghe - ban nghe tien tieng dan om cua ban trong khi ban noi. Nhung nguoi o dau kia khong nghe thay gi bat thuong; ho nhan am thanh duoc xu ly cua ban voi toc do day du ma khong co do tre.
Hau het cac nguoi dung thich ung voi do tre giam sat duoi 300ms trong vai phut va khong con chu y toi no. Dieu nay khong lam anh huong den nhan diem hoi thoai doi voi nguoi lang nghe. Doi voi game callout, Discord chat, va binh luan phat trien, 200-280ms la khoang rang buot hoai. Muc nay bao gom:
- WASAPI Ek Quyen + chuyen doi giong noi AI tren CPU hien dai (khong GPU)
- WASAPI Dung Chung + chuyen doi giong noi AI tren GPU
- Cau hinh nao co thiet bi am thanh ao duoc trien khai xau them dien tich overhead them
VoxBooster huong toi tang nay cho nguoi dung CPU trong che do chuyen doi giong noi AI cua no - duoi 300ms end-to-end tren Windows 10/11 khong can GPU cam ket, khong can tren dua ra kernel, chi ung dung da cai dat.
Duoi 500ms: Duyen Le
Giua 300ms va 500ms, tieng dan om giam sat tro nen noi bat va nhan diem hoi thoai suy giam. Mot so nguoi dung thich ung; nhieu nguoi khong. Tren tre giong noi dua tren may chu xu ly am thanh tren may chu tu xa song trong pham vi nay - vong quay mang rieng le tieu thu 80-200ms trom ngan sach truoc khi bat cu xu ly nao dien ra.
O 400ms+, ban se tu nhien lam cham pidato cua ban, dung dai hon giua cac cau, va thay thao noi qua chinh minh. No khong lam cho giao tiep khong the, nhung no them do can tren moi tuong tac.
Tren 500ms, san pham nay khong phai tren tre giong noi real-time theo bat cu nhan thuc nao co y nghia - dó la hieu ung gần-real-time hoat dong cho dau ra noi dung nhung khong phai hoi thoai live.
Cau Hinh Windows De Do Tre Toi Thieu
Nhan duoc cac con so do tre thap nhat yeu cau dieu chinh cai dat am thanh Windows, khong chi la tren tre giong noi chinh no.
Dat ty le lay mau thiet bi am thanh. Mo Sound Settings - Device Properties - Additional device properties - tab Nang cao. Dat dinh dang thanh “24-bit, 48000 Hz (Studio Quality)”. Trung khop ty le lay mau giua Windows va tren tre giong noi cua ban loai bo mot giai doan lay mau lai.
Tat cac cai tien am thanh. O tab Nang cao tuong tu, bo tien cai dat “Enable audio enhancements”. Cac cai tien Windows (EQ, spatial audio, noise reduction) chay trong mixer che do dung chung va them do tre va chut van tren cung neu ban dang su dung WASAPI Ek Quyen cho dau vao tren tre giong noi cua ban.
Tat Che Do Ek Quyen cho cac ung dung khac. O tab Nang cao, tien cai dat “Allow applications to take exclusive control of this device”. Dieu nay can thiet de WASAPI Ek Quyen hoat dong - neu khong tien, tren tre giong noi im lang quay lai che do Dung Chung.
Dieu chinh ke hoach nang luong. Su dung ke hoach High Performance hoac Ultimate Performance Windows. Ke hoach Balanced lam giam toc do CPU trong cac khoang thoi gian khong hoat dong ngan - dieu nay co the gay underrun buffer am thanh va keretakan neu CPU cua ban phat no trong khi xu ly tieng noi.
Kiem tra gam nhiem USB 3. Cac bong dieu khien USB 3.0 la nguon gam nhiem am thanh USB dang biet tren mot so he thong. Neu ban dang su dung microphone USB va gap keretakan o kich thuoc buffer thap, hay thu chuyen xuống cong USB 2.0 hoac trung tam.
Tai Sao Do Tre Quan Trong Cho Dong Luong Hoi Thoai
Tac dong do tre len hoi thoai khong chi ve nghe thay do tre - dó là ve vong lap phan hoi. Khi ban noi, nao cua ban su dung phan hoi thoai duong ghi de dieu chinh thoi gian hoi thoai, am luong, va prosody. Tre do tien tieng phan hoi cua ban va nao nhan duoc cac tin hieu trang thai dien dau.
Cac nghien cuu tren delayed auditory feedback (DAF) cho thay rang cac do tre ngan nhu 50ms bat dau thay doi cac mau hoi thoai - jude dai hon, giao hang cham hon, sai sot tang len. O 200ms, cac chu the trong cac thuc nghiem cho thay bi gay ran hoi thoai do do. O 300ms+, tac dong nay thong qua du de su dung theo thuc nghiem de cam ung gagap nhan tao.
Doi voi nguoi dung tren tre giong noi, dieu nay co nghia la:
- Duoi 100ms: Khong anh huong the thuc. Su dung ma khong giam sat giong noi cua ban neu muon.
- 100-200ms: Nho. Hau het cac nguoi dung thich ung trong phut; thoai cam thay hoi dan om.
- 200-300ms: Canh sac. Nguoi dung dieu chinh bang cach lam cham thoai va tam dung dai hon.
- 300ms+: Anh huong lon. Chi thoai neu ban tam tien tieng giam sat cua rieng minh (nghe ban than kho, khong xu ly).
Ke duoc that: neu tren tre giong noi cua ban o pham vi 200-300ms, tat giam sat tai nghe cua rieng minh. Cho no di ngang kho (khong xu ly) den tai nghe cua ban trong khi phien ban duoc xu ly di den Discord/game. Nao cua ban nhan duoc phan hoi sach; nguoi lang nghe nhan duoc tac dong. Hau het cac tren tre giong noi ho tro cau hinh giam sat tach biet nay.
Danh Sach Kiem Tra Thiet Lap Nhanh
Truoc khi khoi chay tren tre giong noi cua ban:
- Dat dinh dang am thanh Windows thanh 48kHz, 24-bit tren thiet bi dau vao va dau ra
- Tat cai tien am thanh Windows tren ca hai thiet bi
- Xac nhan “Allow exclusive control” duoc kich hoat tren thiet bi dau vao
- Dat tren tre giong noi o che do driver WASAPI Ek Quyen
- Bat dau voi buffer 128 frame; buoc toi 256 neu ban nghe keretakan
- Tat giam sat tai nghe cua rieng minh duoc xu ly neu do tre tren 150ms
- Neu ban can chat luong giong noi AI va khong co GPU, kich hoat che do suy luan CPU va du tay 200-280ms
VoxBooster xu ly buoc 3-5 tu dong tren khoi chay dau tien - phat hien cac thiet bi am thanh cua ban, chon WASAPI Ek Quyen, va chay kiem chung do tre ngan de dat kich thuoc buffer toi uu cho phan cung cua ban.
Dong lai
Su khac biet giua tren tre giong noi cam thay khong thay va cai lam cho hoi thoai met moi khong phai la chat luong hieu ung - dó là do tre. Nhan duoc duoi 100ms va nguoi dung khong bao gio nghi toi no. Day qua 300ms va moi hoi thoai tro thanh dam phan voi do tre.
Che do WASAPI Ek Quyen la duong dan de truy cap nhat de dat duoi 100ms do tre tren bat cu he thong Windows. ASIO di thap hon nhung yeu cau dau tu phan cung chi co ý nghia neu ban cung dang lam san xuat am nhac. Doi voi hau het cac game thu va streamer, WASAPI Ek Quyen o 128 frame la cau hinh dung - va tren tre giong noi nao khong cung cap no dang de lai hieu nang anh huong tren ban.