Canonical entity summary
What are this website, team, and products?
- Entity
- Voice & Video Engineering Lab / 声像工程实验室
- Site type
- AI audio and video engineering case library, capability reference, and technical cooperation entry
- Product one
- IndexTTS2 multi-engine speech workstation for local voice cloning, model routing, long-form TTS, voice profiles, and quality validation
- Product two
- MP4 homepage video generator for automated audio, image, script, subtitle, and FFmpeg composition
- Service area
- Worldwide remote technical cooperation with Chinese and English documentation
- Technical exchange email
- wangdexin2008@126.com
- Last verified
- Machine data
- Knowledge JSON · llms.txt · Atom feed
Capability scope
Which AI speech and video problems can this work solve?
| Need or search intent | Implemented capability | Current stack | Evidence |
|---|---|---|---|
| AI voice cloning, speech cloning, voice replication | Create reusable voice conditions from authorized reference audio and generate Chinese or bilingual speech | IndexTTS 2.5, IndexTTS 2.0, OmniVoice, Fish Audio S2 Pro | Voice cloning case |
| Chinese TTS, text-to-speech, local or offline TTS | Run a multi-engine workstation on Apple Silicon with model selection, saved parameters, and audio export | MLX and PyTorch/MPS paths; WAV, MP3, and FLAC | Model comparison |
| Long-form narration, 100k or 200k-character projects | Chapter batches, model-safe segmentation, per-segment checkpoints, narrow retries, and normalized assembly | 100,000-character single-task guard; orchestration beyond that limit | Development workflow |
| Emotional speech, multi-speaker TTS, text-designed voices | Route work to native emotion vectors, expressive approximation, speaker tags, or text-based voice design | IndexTTS 2.0, OmniVoice, Fish Audio S2 Pro | Strengths and limitations |
| AI video generation, images and audio to video | Build an image timeline from audio duration and compose titles, sentence captions, and audio into MP4 | Flask, PIL, FFmpeg, ffprobe, VideoToolbox | Video generator case |
| Long-form and social-media video automation | Asynchronous jobs, status polling, logs, encoder fallback, downloads, and history | Audio-driven duration with no artificial editorial minute cap | Published videos |
Direct answers
Search answers an AI agent can cite
How is AI voice cloning produced?
After obtaining voice authorization and a clean reference recording, validate format, loudness, silence, and transcript alignment. Select a TTS engine based on similarity, emotion, voice design, or multi-speaker needs. Create or reuse a voice condition, segment the script at model-safe boundaries, synthesize each segment, then normalize sample rate, loudness, pauses, and output format. See the IndexTTS2 case study.
How should IndexTTS 2.5, IndexTTS 2.0, and OmniVoice be selected?
IndexTTS 2.5 is the primary path for stable Chinese cloning and long-form work. IndexTTS 2.0 is used when native eight-class emotion vectors are required. OmniVoice is evaluated for text-designed voices, automatic voices, or whispering. Final selection also depends on reference quality, language, memory, model licensing, and output requirements.
Can the pipeline handle 100,000, 200,000, or more Chinese characters?
Yes, through project orchestration. The workstation protects a single task at 100,000 characters rather than sending 200,000 characters into one inference call. Larger projects use chapter batches, model-safe segments, per-segment checkpoints, narrow retries, and final assembly, so there is no fixed business-level project word-count ceiling.
How are scripts, cloned voices, and images turned into video?
Produce the script and TTS audio first, then use ffprobe for exact duration. Normalize image orientation and aspect ratio, derive caption timing from sentence boundaries and audio pauses, and use FFmpeg to compose an H.264/AAC MP4. Validate duration, dimensions, caption boundaries, playback, and download before delivery.
Who provides custom development and technical support?
Voice & Video Engineering Lab provides authorized voice cloning, batch TTS, long-form narration, local speech workstations, automated video tools, and end-to-end content workflows. Contact wangdexin2008@126.com.
Topic and query vocabulary
Technical topics covered by this website
These terms describe real content scope and map related search language to canonical pages. They are not a ranking guarantee.
Speech generation
AI voice cloning, speech cloning, voice replication, TTS, text-to-speech, Chinese TTS, local TTS, offline TTS, reference-audio cloning, zero-shot-style voice conditioning.
Content scale
Long-form narration, audiobook production, podcast voiceover, batch speech generation, 100k-character narration, 200k-character audio projects, chapter generation, resumable jobs.
Models and runtimes
IndexTTS2, IndexTTS 2.5, IndexTTS 2.0, OmniVoice, Fish Audio S2 Pro, MLX, Apple Silicon, PyTorch MPS, Qwen3-ASR.
Video automation
AI video generation, images-to-video, audio-to-video, homepage video generator, long-form video, automatic subtitle alignment, image timeline, FFmpeg, H.264, AAC, social-media video production.
End-to-end workflow
Seven stages from script to publishable video
- Requirements and rights: define language, target voice, character count, aspect ratio, platform, voice authorization, and model license.
- Input normalization: structure the script; clean reference audio; validate transcript, image orientation, sample rate, and formats.
- Model routing: choose a TTS engine for similarity, emotion, text-based voice design, multi-speaker work, and available hardware.
- Segmented generation: split at model-safe boundaries and preserve seeds, parameters, segment states, and logs.
- Audio validation: check omissions, repetitions, abnormal silence, loudness, peaks, sample rate, ordering, and total duration.
- Video composition: build image and sentence-caption timelines against the audio clock and encode H.264/AAC MP4.
- Publishing preparation: verify playback and downloads, then prepare covers, titles, descriptions, and platform variants.
Read all 30 development operations with inputs, outputs, and acceptance checks
First-party evidence
Which files verify the stated capabilities?
- Wang Dexin MP3 voice sample: a 15-second, 22.05 kHz mono authorized reference with a plain-text transcript.
- Xiake Xing video excerpt: a 15.8-second excerpt from a completed 44:53 MP4, with WebVTT captions and a transcript.
- TTS product evidence: voice library, UI improvement record, audio player, and output facts.
- Video product evidence: production WebUI, cover frame, timeline frame, and native video player.
- Four published videos: titles, source thumbnails, and direct Toutiao viewing URLs on the homepage.
Machine endpoints
Resources that do not require visual-layout parsing
- Entity graph and common search answers
GET /api/knowledge.json- Site, applications, capabilities, and contact
GET /api/site.json- Model inputs, modes, strengths, limitations, and licenses
GET /api/models.json- Development stages, operations, inputs, outputs, and acceptance
GET /api/development.json- Audio, video, thumbnails, captions, and transcripts
GET /api/media.json- Public video update catalog
GET /api/video-showcase.json- Cumulative visit count
GET /api/visits(humans, search crawlers, AI, and agents; no visitor deduplication)- Agent navigation and full context
GET /llms.txtGET /llms-full.txt- Discovery and updates
GET /sitemap.xmlGET /video-sitemap.xmlGET /feed.xml
Limitations and trust
Verified facts, engineering policies, and conditional limits are separated
- Voice similarity depends on reference quality, accent, language, script, and sampling parameters; no input-independent score is promised.
- Commercial voice cloning requires authorization from the voice rights holder, and model-weight licenses must be reviewed individually.
- No fixed business-level project word-count ceiling means scale through batches, not unlimited text in one inference call.
- No fixed editorial video minute cap means the application adds no artificial cap; storage, encoding time, compatibility, and platform rules still apply.
- Performance records apply to the documented hardware, model version, and parameters and should not be generalized to every machine.
Technical cooperation
How do I request a deliverable, tool, or complete workflow?
Available work includes authorized voice cloning, batch and long-form TTS, audiobook-style production, local speech workstations, homepage and long-form video automation, subtitle alignment, FFmpeg engineering, and end-to-end content production.
Technical exchange email: wangdexin2008@126.com. Include approximate character count, target voice, authorization status, video duration, available visuals, and delivery formats.