Skip to content
Date and source

Dates follow the news timeline in Beijing time and include both endpoints.

Oct 21Wed
  1. mlx-audio · 开源创作项目发布18

    mlx-audio v0.6.1 released: fixes Ming prompt contract and adds Gradium Phonon TTS

    mlx-audio released v0.6.1, fixing the prompt contract for Ming dense and supporting Ming-3.0-tts-0.5B, while adding the Gradium Phonon TTS model. This version also fixes server-side STT uploads to use writable temporary files, and fixes a crash in irodori-tts where speaker_kv_scale rolled back under independent CFG and dual-conditioning settings.

Oct 20Tue
  1. TV Technology22

    From tape to intelligent digital archives: the technological evolution of audiovisual preservation

    Audiovisual archive preservation is shifting from reliance on physical media such as LTO tape to policy-driven digital repositories; LTO-10 native capacity has reached 30TB and 40TB, but generational migration and compatibility remain core risks. High-end scanners such as the Lasergraphics Director 65 can scan up to 13.5K, while routine preservation workflows mostly use 4K. The real transformation after digitization lies in ingesting preservation masters directly into object storage, with software automatically enforcing integrity, retention, and replication policies.

  2. 雷峰网62

    580 days of covert battle, iQIYI's AI journey through the gauntlet

    Leiphone reviews iQIYI's 580-day AI implementation process since company-wide AI training in March 2023, covering applications such as the Script Workshop, Image Workshop, virtual production, and Taodou World. The Script Workshop has cumulatively provided accelerated evaluation for 5,000 novels and scripts; the April 2025 version integrated domestic large models such as DeepSeek, and can complete evaluation of a 500,000-character script in under 30 minutes. The price of the virtual production system has dropped from 160,000 yuan/day to a minimum of 35,000 yuan/day starting June 1 this year, and the number of projects using virtual production internally has increased by 50% year over year.

  3. IT之家68

    Ant Ling opens source Ming-3.0 speech series models, covering speech synthesis, audio understanding and dialogue, and speech recognition

    Ant Ling has open-sourced the Ming-3.0 speech series, including three series—speech synthesis, audio understanding and dialogue, and speech recognition—with five models in total, which developers can select by task or combine into workflows. Ming-3.0-tts's Voice Design supports describing the target timbre in text, and can also set the timbre with reference audio; director mode can specify how lines are performed, emotional changes, stressed words, and pacing.

    Why it matters: Ming-3.0 splits speech synthesis, audio understanding, and recognition into five models, among which director mode can describe the performance style of lines in text.

  4. CGPress72

    Mirelo releases SFX 2.0 video-to-sound-effect model

    Berlin-based AI research team Mirelo released the video-to-sound-effect model SFX 2.0, which can modify a video's audio without regenerating the visuals, preserve dialogue, remove unspecified music, and generate sound effects, Foley and ambience synchronized with action and consistent across cuts.

    Why it matters: The video-to-sound-effect model has been updated to redo the audio track without changing the visuals, and has been integrated into editing software and third-party platforms.

  5. mlx-audio · 开源创作项目发布22

    mlx-audio v0.6.0 released: adds support for Fish Audio Drama 3, Phonon-2 and MiMo-V2.5-ASR

    mlx-audio released v0.6.0, adding support for the Fish Audio Drama 3 speech model, Phonon-2 speech recognition and MiMo-V2.5-ASR STT, plus Granite Speech 4.1 Plus. This version improves Breeze-TTS and VibeVoice generation performance, and fixes Qwen3-ASR streaming timestamps, Qwen3-TTS prefill, and the fallback when Metal is unavailable.

  6. ComeCut · 开源创作项目发布42

    ComeCut v0.2.1 releases its first Windows desktop version, about 30MB

    ComeCut v0.2.1 released its first Windows desktop version, only about 30MB in size, with fast startup and low resource usage. It adds a multilingual interface, automatic dubbing (supporting Chinese, English, Korean and Japanese, requiring the corresponding TTS), and auto-save when exiting the editing editor, and fixes the issue of audio being lost after separation, while displaying vocals and background audio in different colors.

  7. IT之家58

    Japan Actors Union releases AI voice usage guidelines; Dragon Ball Goku voice actor Masako Nozawa angrily criticizes voice misappropriation

    The Japan Actors Union (JAU) has released the first version of its generative AI usage guidelines, applicable to all audio generated by replicating actors' voices with AI. It requires obtaining the person's consent at every step of the replication process; AI companies must obtain permission to use actors' voices for training; authorized generated speech is regarded as a type of "performance," and companies must pay remuneration or royalties.

Oct 19Mon
  1. mlx-audio · 开源创作项目发布33

    mlx-audio v0.6.0 released

    mlx-audio releases v0.6.0, adding support for Fish Audio Drama 3, Phonon-2 speech recognition, MiMo-V2.5-ASR, and Granite Speech 4.1 Plus. It also fixes multiple issues with Qwen3-TTS/ASR, Whisper, and others, and optimizes Breeze-TTS generation and VibeVoice performance.

  2. IT之家38

    NPCTalk brings real-time, script-free voice dialogue to GTA 5 NPCs

    The open-source AI mod NPCTalk brings script-free real-time dialogue to GTA 5: players can talk to NPCs via microphone, and NPCs respond with AI-cloned original voices. The system runs on NVIDIA Nemotron ASR, Qwen 3.8 27B (Cerebras platform), Fish Audio Drama 3, and TypeSafe Jev working together, with push-to-talk latency of about 1.15 to 2.2 seconds.

  3. Hugging Face Daily Papers24

    SEAM: A spectral embedding alignment and merging method for low-resource TTS

    SEAM is a zero-audio initialization method that builds embeddings for unseen tokens using known subwords, preserves temporal phoneme order through spectral filtering, and aligns new embeddings to the pretrained conditioning manifold using closed-form ridge regression. Across three languages and two TTS architectures, SEAM approaches or matches baselines with up to 4x less training data, reducing sample complexity by 75% compared with random initialization and mean pooling. The work has been accepted to EMNLP 2026 Findings.

Oct 18Sun
  1. 开源中国60

    AIGCPanel v2.6.0 supports generating digital human avatars directly from photos, with adjustable speech synthesis parameters

    AIGCPanel v2.6.0 released, adding an "Image Avatar" entry: upload a high-definition front-facing selfie to generate a digital human avatar, no longer requiring recording a 10-to-30-second front-facing video. This version also supports fine-tuning speech synthesis parameters during generation and adds a fallback mechanism for task scheduling.

  2. Production Expert38

    Session Loops releases free open-source AI pitch-correction plugin PitchNet

    Session Loops has launched PitchNet, a free open-source AI vocal pitch-correction plugin that uses AI resynthesis to rebuild the corrected vocal rather than directly distorting the original waveform, and can detect vocals as editable note regions. The plugin supports VST3, AU, AAX and ARA; the Mac version is a universal binary for Apple Silicon and Intel; the official recommendation is to use a discrete GPU, though it can also run at reduced speed on integrated graphics.

  3. z-ab · 开源短剧与创作项目分享62

    Open-Source Claude Skill Lets the Browser Draw Hand-Drawn Storybook Animation in Real Time, with Voiceover and Music in One Go

    An open-source Claude Skill called hand-drawn-storybook-animation-skill can draw hand-drawn watercolor storybook animation in real time in the browser, with the image swept out stroke by stroke, including storyboards and page-turn transitions, rather than pre-rendered video.

Oct 17Sat
  1. FastVideo 视频模型 · 模型发布60

    FastVideo releases FastH3-QAD-v2 video-audio model

    FastVideo releases FastH3-QAD-v2, an 8-step video and audio model trained with NVFP4 weights and quantization-aware distillation (QAD), runnable on a single Blackwell GPU (RTX 5090, RTX PRO 6000) or DGX Spark.

    Why it matters: The 8-step video and audio model lands on a single Blackwell GPU with NVFP4 quantized weights, and error data versus the BF16 teacher model is provided.

  2. Hugging Face Daily Papers38

    UniAVC: unified audio-video creation via progressive training and asymmetric scheduling

    UniAVC uses a single set of model weights to achieve text-conditioned joint generation, bidirectional cross-modal generation, and instruction-guided editing without task-specific adapters. It first trains each branch on unimodal data to full generation and editing capability, then uses paired data for cross-modal correspondence; a modality-decoupled scheduler independently transforms noise schedules from a shared timestep, and directly supervises the asymmetric configuration with unimodal clean samples. This enables intent-aware editing and post-inference self-refinement, regenerating only the unsatisfactory side.

  3. Reddit comfyui · 创作与工作流分享12

    minimax H3 Voice Prompts with Audio Reference Fail, Cannot Change Volume or Timbre

    A user reported that in the minimax H3 reference-image-to-video workflow, after inputting about 20 seconds of normal speech as an audio reference, voice prompts produced no change at all, and all outputs had the same volume and timbre. Attempts to use annotations such as says and whispers quietly were ineffective, and changes such as shouting, whispering, or an excited tone could not be achieved.

  4. ElevenLabs 产品与研究62

    ElevenLabs launches Instant Soundtracks in ElevenCreative

    ElevenLabs Launches Instant Soundtracks in ElevenCreative; After Uploading a Video, the Model Generates an Original Score and Analyzes Sound Effect Cue Points in the Footage, Automatically Placing Sound Effects on the Timeline at the Corresponding Timestamps.

    Why it matters: ElevenLabs merges score and sound effect generation into the video workflow, which can be used to evaluate the sound production stage of short dramas and ads.

Oct 16Fri
  1. 中广互联 · 短剧22

    Broadcast Network Launches Pre-Reorganization; China Broadcasting Network Shanxi Company Advances "One Network for the Whole Province" Integration

    Broadcast Network launched pre-reorganization after creditor Xi'an Guanghua Communication applied on the grounds of 5.1133 million yuan in matured debt unpaid; the Xi'an Intermediate People's Court has accepted the case and designated a liquidation group as interim administrator, while publicly recruiting reorganization investors, with registration closing at 18:00 on November 8, 2026 and a 10 million yuan deposit required. China Broadcasting Network Shanxi Company held a "Hundred-Day Sprint, Win the Year" progress meeting on October 10, calling for accelerating the deep integration of "one network for the whole province" and improving the three-tier business data reporting mechanism.

  2. 我爱计算机视觉 · 公开公众号 RSS31

    HKUST × Tencent Video propose Adaptive Reward Routing: audio-video diffusion model post-training no longer fixes reward routing in advance

    MMLab@HKUST of HKUST, Tencent Video, and the University of Hong Kong propose Adaptive Reward Routing, which, on the DiffusionNFT framework, recomputes multi-reward routing for audio-video diffusion models in real time during training rather than fixing it once from the base model.

Oct 15Thu
  1. note · kamiryu_ai · 图像视频创作分享36

    Making a 15-second coffee product ad with AI video: generated with Higgsfield, edited with Claude Code, but the final product bag couldn't be fixed correctly by AI

    A creator used Higgsfield to generate a 15-second vertical 720p coffee drip bag ad in one go, consuming 68 credits, but the product bag appearing in the last approximately 2.4 seconds was replaced by AI with a large stand-up pouch, and the text on the bag became meaningless letters.

  2. note · toshia_fuji · 图像视频创作分享22

    Generative AI news roundup: DigUp, Fish Audio Drama 3, Mirelo SFX 2.0, OmniHuman 2.0 and other new tools released

    Fish Audio released the voice AI "Drama 3," supporting up to 16 speakers, emotional performance, and voice editing, with the model data and code open-sourced. Mirelo SFX 2.0 can automatically generate synchronized sound effects for videos and remove unwanted music, processing about 10-minute clips per run. BytePlus's OmniHuman 2.0 supports generating talking and singing videos up to 300 seconds long from a single image plus audio, and can be called via API.

  3. Production Expert39

    Why audio professionals need agentic AI, not generative AI

    DaVinci Resolve 21 has integrated agentic AI; Claude or ChatGPT can control the software via API with prompts. In testing, it automatically synced four clips from two cameras without timecode and one audio track without timecode on the timeline, generated a transcript, picked out the best two minutes, and performed loudness normalization to YouTube standards.

  4. note · koheii_miaa · 图像视频创作分享65

    How an AI musician uses a single Mac to produce an MV with lyric sync and lip-sync, 17 minutes per track

    The author published a workflow for locally generating MVs on a Mac: input a song and cover image, output a 30-second video with synchronized lyrics and a cover character singing, in both 16:9 and vertical formats, approximately 17 minutes per track, without using cloud-based paid generation.

    Why it matters: The author published the complete local workflow from vocal separation to lyric sync with time breakdowns, allowing for an assessment of the feasibility of producing MVs on a single machine.