AnifLive-TTS v1.5.0-beta

AnifLive-TTS v1.5.0-beta

English

Highlights

• Added the official Python SDK with synchronous Client and asynchronous AsyncClient, backed by a shared Rust transport core.
• Typed service discovery, model activation, complete WAV synthesis, incremental PCM streaming, cancellation, and speech sessions.
• Session lookup, ordered segments, idempotency, flush, and cancellation follow the existing server API.
• Lightweight client installation with no local CUDA, TensorRT, or model requirement. Server commands use the runtime extra.

SDK

Install the supplied wheel matching your platform, then connect to a running AnifLive-TTS service. See the SDK manual for download, installation, sync/async examples, expressions, sessions, and typed errors.

The package version is 1.5.0b1. Linux x64 wheels target CPython 3.10–3.12 with a manylinux_2_28 baseline. Windows x64 and macOS arm64 wheel workflows are prepared; their platform qualification is pending.

Runtime, WebUI And Studio

The existing TensorRT runtime, WebUI, and Studio remain available. This beta adds a client SDK; it does not embed inference in client applications. Docker builds compile the client extension in a separate builder stage.

繁體中文

主要更新

• 新增官方 Python SDK，提供同步 Client 與非同步 AsyncClient，共用 Rust 傳輸核心。
• 型別化服務探索、模型啟用、完整 WAV 合成、增量 PCM 串流、取消及 Speech Session。
• Session 查詢、順序片段、冪等追加、flush 與取消沿用現有伺服器 API。
• 輕量用戶端不需要本機 CUDA、TensorRT 或模型；伺服器指令使用 runtime extra。

SDK

安裝符合平台的隨附 wheel，再連接已啟動的 AnifLive-TTS 服務。下載、安裝、同步／非同步範例、情感設定、session 與型別化錯誤請參閱 SDK 使用手冊。

套件版本為 1.5.0b1。Linux x64 wheel 以 manylinux_2_28 為基準，目標為 CPython 3.10–3.12。Windows x64 與 macOS arm64 的 wheel 工作流程已備妥，平台驗證尚待執行。

Runtime、WebUI 與 Studio

既有 TensorRT runtime、WebUI 與 Studio 繼續並存。本次 Beta 新增用戶端 SDK，不會把推理嵌入用戶端應用程式。Docker 透過獨立建置階段編譯 client extension。

AnifLive-TTS v1.4.0

AnifLive-TTS v1.4.0

English

AnifLive-TTS v1.4 adds the local Studio training workstation while preserving
the V2ProPlus FP16 TensorRT 11 inference path.

Highlights

• Adds dataset preparation, review, training, checkpoint selection, conversion,
  evaluation and qualified model installation in Studio.
• Uses dataset-duration training recommendations and human listening to guide
  checkpoint selection.
• Supports epoch-boundary pause, real checkpoint continuation and managed GPU
  handoff between training/evaluation workers and inference.
• Stores semantic sampling with each newly exported model package and preserves
  the existing sampling behavior of older packages.
• Adds a Docker Studio launcher and supports named model folders on first use.

Studio and WebUI

Studio and WebUI coexist. Studio provides the dataset, training, conversion,
evaluation and model-management workstation; run run_studio_docker.bat and
open 127.0.0.1:9891. WebUI remains the lightweight model/expression/speech
interface; start the API, run run_webui.bat and open 127.0.0.1:9890.
See the Studio guide for the complete workflow.

Runtime

• API version: 1.4.0
• Backend: TensorRT 11, FP16
• Languages: zh, yue, en, ja, ko
• Docker profiles: cu128 by default; cu126 is prepared as a compatibility
  build and has not been GPU-tested on the current acceptance machine.

Original AnifLive-TTS code remains licensed under PolyForm Noncommercial
1.0.0. Commercial use requires a separate written Commercial License from
Hiruynk; upstream and third-party components retain their respective licenses.

繁體中文

AnifLive-TTS v1.4 加入本機 Studio 訓練工作站，並保留既有 V2ProPlus FP16
TensorRT 11 推理路徑。

主要更新

• Studio 支援資料集準備、審核、訓練、檢查點選擇、轉換、評估及合格模型安裝。
• 根據資料集時長提供訓練建議，配合人工盲聽選擇檢查點。
• 支援在訓練回合邊界暫停、從真實檢查點續訓，以及訓練／評估與推理之間的 GPU 交接。
• 新匯出的模型套件會記錄語意採樣策略，舊套件保留原有採樣行為。
• 新增 Docker Studio 啟動器，首次使用亦支援具名模型資料夾。

Studio 與 WebUI

Studio 與 WebUI 並存。Studio 提供資料集、訓練、轉換、評估及模型管理工作站，
執行 run_studio_docker.bat 後開啟 127.0.0.1:9891。
WebUI 保留模型選擇、情感控制及語音播放的輕量介面；先啟動 API，再執行
run_webui.bat，位址為 127.0.0.1:9890。
完整流程請見 Studio 使用手冊。

執行環境

• API 版本：1.4.0
• 後端：TensorRT 11、FP16
• 語言：zh、yue、en、ja、ko
• Docker 配置：預設 cu128；另準備 cu126 相容建置，未在本次驗收電腦執行 GPU 測試。

AnifLive-TTS 原創程式碼繼續採用 PolyForm Noncommercial 1.0.0；商業使用須
另行取得 Hiruynk 的書面商業授權，上游與第三方組件維持各自授權。

AnifLive-TTS v1.3.0

AnifLive-TTS v1.3.0

English

AnifLive-TTS v1.3 adds package-curated expression control and the local WebUI
while preserving the validated V2ProPlus FP16 TensorRT 11 inference path.

Highlights

• Adds global and per-segment symbolic expression requests with discrete
  intensity selection.
• Prepares every package-owned expression reference at model activation and
  keeps the neural conditioning resident on the GPU.
• Adds the generic model import-expressions administration command; it uses
  model-package metadata and does not branch on a character or model name.
• Includes a five-language local WebUI with model switching, natural-language
  expression selection, streaming playback, cancellation, and live metrics.
• Preserves neutral complete WAV, streaming PCM, and semantic output exactly
  against the immutable v1.2 baseline for both qualified packages in all five
  languages.
• Retains the existing v1 API, GPT-SoVITS flat adapter, OpenAI-compatible
  adapter, schema-1 model packages, and Transformer/TensorRT acoustic runtime.

Quality and architecture decision

Both qualified V2ProPlus packages passed the Japanese full-expression matrix,
the five-language long/short matrix, real TensorRT API acceptance, and the
independent blind-listening gate. v1.3 deliberately keeps the v1.2 Transformer
semantic decoder after hard-token acoustic lookahead, recurrent speculative
decoding, SemanticPiece/EABPE, FastStart, and the evaluated D16 block-diffusion
recipe failed their respective release gates.

The implementation and expression policy are package-driven. The runtime has
no Miku- or Roxy-specific branch, and one process continues to load only one
active model package at a time.

WebUI

Start the API, then run run_webui.bat. The local WebUI binds to
127.0.0.1:9890 and provides model switching, segmented expression control,
streaming playback, and live latency metrics.

Runtime

• API version: 1.3.0
• Backend: TensorRT 11, FP16
• Languages: zh, yue, en, ja, ko
• Docker profiles: cu128 by default, with cu126 as a compatibility build

The canonical public performance report uses Roxy only. Miku remains outside
the performance headline while its model-specific long-form streaming behavior
is investigated separately.

Original AnifLive-TTS code remains licensed under PolyForm Noncommercial
1.0.0. Commercial use requires a separate written Commercial License from
Hiruynk; upstream and third-party components retain their respective licenses.

繁體中文

AnifLive-TTS v1.3 在保留既有 V2ProPlus FP16 TensorRT 11 推理路徑的同時，
加入由模型套件管理的情感控制及本機 WebUI。

主要更新

• 支援全句及分段的情感設定，並可選擇離散強度。
• 模型啟用時會準備套件內所有情感參考資料，神經網路條件資料持續保留於 GPU。
• 新增通用的 model import-expressions 管理命令；工具只依賴模型套件資料，
  不會根據角色或模型名稱分支。
• 加入五語本機 WebUI，提供音色切換、自然語言情感選擇、串流播放、取消請求及
  即時數據。
• 兩個完成驗收的模型套件，在五種語言下的中性完整 WAV、串流 PCM 及語意輸出
  均與不可變的 v1.2 基準完全一致。
• 保留既有 v1 API、GPT-SoVITS 舊版相容介面、OpenAI 相容介面、schema-1
  模型套件及 Transformer／TensorRT 聲學推理流程。

音質與架構決定

兩個完成驗收的 V2ProPlus 模型套件均通過日語完整情感矩陣、五語長短句矩陣、
真實 TensorRT API 驗收及獨立盲聽。硬式語意預讀、循環式推測解碼、
SemanticPiece／EABPE、FastStart 及本次 D16 區塊擴散方案未通過各自的發行門檻，
因此 v1.3 刻意保留 v1.2 Transformer 語意解碼器。

情感策略完全由模型套件決定，執行環境沒有 Miku 或 Roxy 專屬分支；每個處理程序
仍然只會載入一個啟用中的模型套件。

WebUI

先啟動 API，再執行 run_webui.bat。本機 WebUI 綁定
127.0.0.1:9890，提供音色切換、分段情感控制、串流播放及即時延遲數據。

執行環境

• API 版本：1.3.0
• 後端：TensorRT 11、FP16
• 語言：zh、yue、en、ja、ko
• Docker 配置：預設使用 cu128，另提供 cu126 相容建置

公開性能主數據只採用 Roxy。Miku 的模型特定長句串流現象仍在獨立調查，
因此不納入性能主數據。

AnifLive-TTS 原創程式碼繼續採用 PolyForm Noncommercial 1.0.0；商業使用須
另行取得 Hiruynk 的書面商業授權，上游與第三方組件維持各自授權。

AnifLive-TTS v1.2.0

AnifLive-TTS v1.2.0

English

AnifLive-TTS v1.2 improves interactive latency and complete-generation speed
while preserving the established FP16 TensorRT 11 acoustic path and v1 API.

Highlights

• Adds fitted GPT engines for each compatible voice package, with persistent
  TensorRT execution contexts, fixed KV buffers, and zero auxiliary streams.
• Keeps semantic sampling on the GPU through a reusable CUDA Graph while
  preserving the existing seed and sampling contract.
• Restricts the 9+8 semantic-token preview to the first text segment; later
  segments retain native full-context refill for stable long-form delivery.
• Retains batched two-step EOS checks and the guarded 25-second warm window.
• Validates the release with interleaved Miku and Roxy model sessions rather
  than reporting a result from one voice package.
• Restores the default repetition penalty to 1.0; the fixed Miku complete-WAV
  regression case is byte-identical to v1.1 at the same seed and settings.

Architecture decision

v1.2 evaluated MTP-4, a 1:1 Transformer/Mamba-2 hybrid, and a combined
Mamba-2 + MTP path. None met AnifLive-TTS's joint latency-and-quality
acceptance criteria, so the production runtime deliberately retains the
1-token/NFE Transformer decoder and focuses on TensorRT runtime optimization.
This reduced keep-alive audible TTFA P50 from 89.296 ms to 77.717 ms
without replacing the validated production architecture.

The standalone TensorRT 11 Mamba-2 IPluginV3 feasibility gate passed; the
rejected result applies to the trained hybrid semantic backend on this
workload, not to Mamba-2 or TensorRT plugin feasibility in general. Detailed
findings are in
docs/research/v1.2-semantic-experiments.md.

Known observation

• Blind long-form listening found slight overlap in both v1.1 and v1.2 Miku
  streaming output. It was not observed with Roxy and is scheduled for
  model-specific investigation after v1.2.

The machine-readable benchmark and gate reports are available in
benchmarks/README_BENCHMARK_SUMMARY.json
and benchmarks/V1_2_RELEASE_GATE.json.

Runtime

• API version: 1.2.0
• Backend: TensorRT 11, FP16
• Languages: zh, yue, en, ja, ko
• Release container tag: ghcr.io/hiruynk/aniflive-tts:1.2.0-cu128
• Compatibility profile: cu126; publish its tag only after release-workflow
  and target-host validation

The v1 API and schema-1 model packages remain compatible. Model packages and
TensorRT engines remain external runtime assets and must match the target GPU,
CUDA runtime, and TensorRT fingerprint.

Original AnifLive-TTS code remains licensed under PolyForm Noncommercial
1.0.0. Commercial use requires a separate written Commercial License from
Hiruynk; upstream and third-party components retain their respective licenses.

繁體中文

AnifLive-TTS v1.2 在維持既有 FP16 TensorRT 11 聲學路徑及 v1 API 的同時，
進一步降低互動延遲及完整語音生成時間。

主要更新

• 為每個相容音色套件建立針對其尺寸的 GPT 引擎，重用 TensorRT 執行環境與
  固定 KV 緩衝區，並停用輔助串流。
• 以可重用 CUDA Graph 在 GPU 上完成語意採樣，同時保留既有隨機種子與採樣規格。
• 只有第一個文字分段使用 9+8 語意標記預覽；後續分段沿用原生完整上下文補充路徑，
  維持長句穩定性。
• 保留每兩步批次檢查 EOS，以及受保護的 25 秒熱狀態維持。
• 正式數據採用 Miku 與 Roxy 交錯模型輪次，不再只以單一音色套件作為代表。
• 將預設重複懲罰恢復為 1.0；固定 Miku 完整 WAV 回歸案例在相同隨機種子及
  設定下，與 v1.1 完全一致。

架構決定

v1.2 實際評估了 MTP-4、1:1 Transformer／Mamba-2 混合架構，以及
Mamba-2 + MTP 組合路徑。這些方案都未同時通過 AnifLive-TTS 的延遲與音質
驗收門檻，因此正式執行環境刻意保留 1 token/NFE Transformer 解碼器，並集中
最佳化 TensorRT 執行路徑。最終在不更換已驗證正式架構的情況下，將持續連線
可聽 TTFA P50 由 89.296 ms 降至 77.717 ms。

獨立 TensorRT 11 Mamba-2 IPluginV3 可行性驗證已通過；被否決的是目前
測試工作負載下完成訓練的混合語意後端，而不是 Mamba-2 或 TensorRT 外掛本身。
詳細結果見
docs/research/v1.2-semantic-experiments.md。

已知觀察

• 長句盲聽在 v1.1 與 v1.2 的 Miku 串流輸出均發現少量疊音；Roxy 未出現同類
  問題。這項音色特定問題將於 v1.2 之後繼續調查。

可機讀性能與驗收結果見
benchmarks/README_BENCHMARK_SUMMARY.json
及 benchmarks/V1_2_RELEASE_GATE.json。

執行環境

• API 版本：1.2.0
• 後端：TensorRT 11、FP16
• 語言：zh、yue、en、ja、ko
• 發布容器標籤：ghcr.io/hiruynk/aniflive-tts:1.2.0-cu128
• 相容配置：cu126；必須通過發布流程與目標主機驗收後才發布其標籤

v1 API 與 schema-1 模型套件保持相容。模型套件及 TensorRT 引擎繼續作為外部
執行資產保存，並須符合目標 GPU、CUDA 執行環境與 TensorRT fingerprint。

AnifLive-TTS 原創程式碼繼續採用 PolyForm Noncommercial 1.0.0；商業使用須另行
取得 Hiruynk 的書面商業授權，上游與第三方組件則維持各自授權。

AnifLive-TTS v1.1.0

AnifLive-TTS v1.1.0

English

AnifLive-TTS v1.1 focuses on interactive latency and natural long-form delivery while preserving the FP16 TensorRT 11 inference path.

Highlights

• Punctuation-aware inference starts a new model invocation at safe comma, period, semicolon, exclamation-mark, and question-mark boundaries.
• Long-form streaming starts each punctuation-delimited segment after prior PCM is handed to the HTTP stream; it does not wait for client-side playback to finish.
• Adaptive boundary timing removes excessive model-generated trailing silence while retaining natural comma, sentence, and paragraph pauses.
• A guarded 25-second warm-retention window reduces latency after short idle periods without locking GPU clocks or changing power and fan settings.
• All five advertised language frontends are prepared during startup, removing first-request lazy-loading stalls.
• Client disconnects cancel streaming work promptly; concurrent requests receive HTTP 429 instead of waiting in a hidden queue.
• Long semantic sequences are decoded in TensorRT profile-safe chunks without a PyTorch fallback.
• Compatible local V2ProPlus packages can be discovered and activated without changing the synthesis API; the active package is fully unloaded before its replacement loads.
• Checkpoint inspection and package promotion use model metadata and tensor contracts rather than character-specific filenames.

Runtime

• API version: 1.1.0
• Backend: TensorRT 11, FP16
• Languages: zh, yue, en, ja, ko
• Release container tag: ghcr.io/hiruynk/aniflive-tts:1.1.0-cu128
• Provisional compatibility profile: cu126; publish its tag only after release-workflow and target-host validation

Model packages and TensorRT engines remain external runtime assets. TensorRT engines must match the target GPU, CUDA runtime, and TensorRT fingerprint.

Original AnifLive-TTS code remains licensed under PolyForm Noncommercial 1.0.0. Commercial use requires a separate written Commercial License from Hiruynk; upstream and third-party components retain their respective licenses.

繁體中文

AnifLive-TTS v1.1 集中改善互動延遲與長句自然度，同時維持 FP16 TensorRT 11 推理路徑。

主要更新

• 遇到適合分句的逗號、句號、分號、感嘆號與問號時，立即以新的模型調用推理下一小句。
• 上一小句的 PCM 交給 HTTP 串流後便會開始處理下一個標點分段，不會等待用戶端播放完畢。
• 自適應句界處理會移除模型產生的過長尾靜音，同時保留自然的逗號、句末與段落停頓。
• 加入受溫度、GPU 佔用率與請求狀態保護的 25 秒熱啟動維持，不鎖定 GPU 時脈，也不修改功耗或風扇設定。
• 五種語言的文字前端會在啟動時完成準備，避免首次請求出現延遲尖峰。
• 用戶端中斷連線後會及時取消串流工作；並行請求會收到 HTTP 429，不會進入隱藏佇列。
• 過長語意序列會按 TensorRT profile 安全範圍分塊解碼，不會回退至 PyTorch 模型推理。
• 可在不改變語音合成 API 的情況下尋找並切換相容的本機 V2ProPlus 套件；目前套件會先完整卸載，之後才載入替代套件。
• 模型檢查與套件發佈依據模型資料及張量規格，不依賴特定角色檔名。

執行環境

• API 版本：1.1.0
• 後端：TensorRT 11、FP16
• 語言：zh、yue、en、ja、ko
• 發布容器標籤：ghcr.io/hiruynk/aniflive-tts:1.1.0-cu128
• 暫定相容配置：cu126；必須通過發布流程與相容主機驗收後才發布其標籤

模型套件與 TensorRT 引擎繼續作為外部執行資產保存。TensorRT 引擎必須與目標 GPU、CUDA runtime 及 TensorRT fingerprint 相符。

AnifLive-TTS 原創程式碼繼續採用 PolyForm Noncommercial 1.0.0；商業使用須另行取得 Hiruynk 的書面商業授權，上游與第三方組件則維持各自授權。

AnifLive-TTS v1.0.0

AnifLive-TTS v1.0.0

English

Highlights

• The first-party TTS runtime for AnifEngine-Voice, built for low-latency, high-quality multilingual voice cloning.
• Eight neural inference stages run through FP16 TensorRT 11 with no PyTorch model fallback.
• Canonical, legacy GPT-SoVITS, and OpenAI-compatible APIs support Putonghua/Mandarin (zh), Cantonese (yue), English (en), Japanese (ja), and Korean (ko).
• v1 launches with GPT-SoVITS V2ProPlus model packaging and inference. More GPT-SoVITS model generations are planned without changing the public API contract.
• Docker images are provided for CUDA 12.8 and CUDA 12.1. Model packages and runtime caches live in mounted host directories, so recreating a container does not download or rebuild them.

Distribution and Compatibility

• Container tags: ghcr.io/hiruynk/aniflive-tts:1.0.0-cu128 and ghcr.io/hiruynk/aniflive-tts:1.0.0-cu121.
• The source bundle is generated from the tagged Git commit. Image-derived SPDX SBOMs for both CUDA variants are attached to this release.
• The API request path performs no dependency installation, engine building, checkpoint loading, or network download.
• CUDA 12.1 engines pass build and load validation. RTX 5070 Ti requires CUDA 12.8, so CUDA 12.1 GPU end-to-end inference is not claimed on this machine.
• Original AnifLive-TTS code is licensed under PolyForm Noncommercial 1.0.0. Commercial use requires a separate written Commercial License from AnifEngine. Upstream and third-party components retain their respective licenses; see LICENSING.md, THIRD_PARTY_NOTICES.md, and the attached SBOMs.

繁體中文

重點

• AnifEngine-Voice 的第一方 TTS，專注於低延遲、高音質的多語言語音複製。
• 八個神經網絡推理階段均使用 FP16 TensorRT 11 執行，不會回退至 PyTorch 模型推理。
• 標準 API、GPT-SoVITS 舊版相容 API 及 OpenAI 相容 API 支援普通話（zh）、廣東話（yue）、英文（en）、日文（ja）及韓文（ko）。
• v1 首發支援 GPT-SoVITS V2ProPlus 模型封裝與推理，未來會在維持公開 API 規格的前提下支援更多 GPT-SoVITS 模型世代。
• 提供 CUDA 12.8 與 CUDA 12.1 Docker 映像。模型套件與快取保存在掛載的主機目錄，重新建立容器時毋須再次下載或建置。

發布與相容性

• 容器標籤：ghcr.io/hiruynk/aniflive-tts:1.0.0-cu128 及 ghcr.io/hiruynk/aniflive-tts:1.0.0-cu121。
• 原始碼套件由對應 tag 的 Git commit 產生；兩個 CUDA 版本由實際映像產生的 SPDX SBOM 亦已附於本 release。
• API 請求期間不會安裝依賴、建立引擎、載入 checkpoint 或從網絡下載檔案。
• CUDA 12.1 引擎已通過建置與載入驗證。RTX 5070 Ti 最低需要 CUDA 12.8，因此本機不宣稱已完成 CUDA 12.1 GPU 端到端推理驗證。
• AnifLive-TTS 原創程式碼採用 PolyForm Noncommercial 1.0.0；商業使用須另行取得 AnifEngine 的書面商業授權。上游及第三方組件維持各自授權，詳見 LICENSING.md、THIRD_PARTY_NOTICES.md 與 release 附件中的 SBOM。