A local-first 82M-parameter TTS workspace with 173 voices across 11 languages, five independently managed model packs, streaming, SSML dialogue, OpenAI-compatible APIs, and opt-in MCP tools.
次の場合におすすめ: Fast narration, multilingual applications, Chinese speech, multi-character dialogue, agent speech workflows, first-time local TTS users, and lightweight integrations.
実行環境: Runs on CPU; NVIDIA GPU acceleration is optional. Model packs load only when one of their voices is used.
メモリ要件8 GB is comfortably in range. The official 70-voice baseline peaked at 2,052 MiB process-reserved VRAM; the newer Chinese v1.1 pack was added after that run.
速度 / RTF15.60× real time overall on RTX 5070 Ti across the official 70-voice, 11-language benchmark. The 173-voice catalog still needs a refreshed complete run.
A local TTS workspace with eight models across six languages, a responsive browser UI, OpenAI-compatible and native APIs, six output formats, and sentence-streamed MP3/PCM.
次の場合におすすめ: English, Spanish, French, Chinese, Japanese, or Korean speech with deterministic model selection, practical output controls, and standard client integrations.
実行環境: CPU and NVIDIA GPU modes are supported; the full image includes every core model, while the smaller English-family image includes EN, EN_V2, and EN_NEWEST.
メモリ要件8 GB is the conservative picker tier; a measured minimum is not published yet.
速度 / RTFNo verified RTX 5060 Ti or RTX 5070 Ti real-time factor is published yet.
A local voice studio for 600+ languages with voice design, microphone and reference cloning, reusable profiles, SSML-H dialogue, streaming, and OpenAI-compatible APIs.
次の場合におすすめ: Multilingual products, structured or multi-character speech, expressive voice design, consented voice cloning, and OpenAI-compatible local applications.
実行環境: GPU is recommended; CPU fallback needs roughly 2–7 GB RAM depending on cloning and ASR use. The default model checkpoint is noncommercial and the images contain additional third-party terms.
メモリ要件The v1.0 release's warmed RTX 5070 Ti benchmark peaked at 4,764 MiB whole-device VRAM; 6 GB is the practical floor and 8 GB is comfortable.
速度 / RTFThe current 18-language regression workload measured RTF 0.295–0.303 on RTX 5070 Ti, or about 3.30–3.39× real time.
A local VoxCPM2 workspace for 30-language generation, voice design, reference and transcript-guided cloning, reusable voice profiles, progressive MP3 streaming, and HTTP APIs.
次の場合におすすめ: Creative voices, realistic consented cloning, multilingual output, reusable reference voices, and advanced local speech applications.
実行環境: The standard Nano-vLLM images require a supported NVIDIA GPU on Windows or Linux. Run one VoxCPMTTS model-serving container per GPU.
メモリ要件The measured Nano baseline peaked at 8,491 MiB on RTX 5070 Ti; 12 GB is the conservative picker tier.
速度 / RTFMedian Nano real-time factor was 0.271 on RTX 5070 Ti, 4.19× faster than the measured native backend without an aggregate transcript-fidelity regression.
A current local ASR service for 30-language transcription, bounded live microphone sessions, exact timestamps and subtitles, OpenAI-compatible clients, and opt-in MCP transcription from local files or URLs.
次の場合におすすめ: Transcription, captions, meeting audio, voice interfaces, agent toolchains, and applications already using OpenAI-shaped APIs.
実行環境: NVIDIA GPU workflow with built-in vLLM 0.26; the default 0.6B model is the lighter starting point and the 1.7B model trades memory for capability.
メモリ要件Measured startup: 4,620 MiB (0.6B) and 7,060 MiB (1.7B). The optional forced aligner adds roughly 3.3 GiB when first loaded.
速度 / RTFThe 0.6B model processed 1,320.936 seconds of 30-language audio in 32.376 seconds on an RTX 5070 Ti, or 40.80 times real time, with a 96.10% benchmark score.
A local-first 82M-parameter TTS workspace with 173 voices across 11 languages, five independently managed model packs, streaming, SSML dialogue, OpenAI-compatible APIs, and opt-in MCP tools.
次の場合におすすめ: Fast narration, multilingual applications, Chinese speech, multi-character dialogue, agent speech workflows, first-time local TTS users, and lightweight integrations.
実行環境: Runs on CPU; NVIDIA GPU acceleration is optional. Model packs load only when one of their voices is used.
メモリ要件8 GB is comfortably in range. The official 70-voice baseline peaked at 2,052 MiB process-reserved VRAM; the newer Chinese v1.1 pack was added after that run.
速度 / RTF15.60× real time overall on RTX 5070 Ti across the official 70-voice, 11-language benchmark. The 173-voice catalog still needs a refreshed complete run.
A local TTS workspace with eight models across six languages, a responsive browser UI, OpenAI-compatible and native APIs, six output formats, and sentence-streamed MP3/PCM.
次の場合におすすめ: English, Spanish, French, Chinese, Japanese, or Korean speech with deterministic model selection, practical output controls, and standard client integrations.
実行環境: CPU and NVIDIA GPU modes are supported; the full image includes every core model, while the smaller English-family image includes EN, EN_V2, and EN_NEWEST.
メモリ要件8 GB is the conservative picker tier; a measured minimum is not published yet.
速度 / RTFNo verified RTX 5060 Ti or RTX 5070 Ti real-time factor is published yet.
A local voice studio for 600+ languages with voice design, microphone and reference cloning, reusable profiles, SSML-H dialogue, streaming, and OpenAI-compatible APIs.
次の場合におすすめ: Multilingual products, structured or multi-character speech, expressive voice design, consented voice cloning, and OpenAI-compatible local applications.
実行環境: GPU is recommended; CPU fallback needs roughly 2–7 GB RAM depending on cloning and ASR use. The default model checkpoint is noncommercial and the images contain additional third-party terms.
メモリ要件The v1.0 release's warmed RTX 5070 Ti benchmark peaked at 4,764 MiB whole-device VRAM; 6 GB is the practical floor and 8 GB is comfortable.
速度 / RTFThe current 18-language regression workload measured RTF 0.295–0.303 on RTX 5070 Ti, or about 3.30–3.39× real time.
A local VoxCPM2 workspace for 30-language generation, voice design, reference and transcript-guided cloning, reusable voice profiles, progressive MP3 streaming, and HTTP APIs.
次の場合におすすめ: Creative voices, realistic consented cloning, multilingual output, reusable reference voices, and advanced local speech applications.
実行環境: The standard Nano-vLLM images require a supported NVIDIA GPU on Windows or Linux. Run one VoxCPMTTS model-serving container per GPU.
メモリ要件The measured Nano baseline peaked at 8,491 MiB on RTX 5070 Ti; 12 GB is the conservative picker tier.
速度 / RTFMedian Nano real-time factor was 0.271 on RTX 5070 Ti, 4.19× faster than the measured native backend without an aggregate transcript-fidelity regression.
A current local ASR service for 30-language transcription, bounded live microphone sessions, exact timestamps and subtitles, OpenAI-compatible clients, and opt-in MCP transcription from local files or URLs.
次の場合におすすめ: Transcription, captions, meeting audio, voice interfaces, agent toolchains, and applications already using OpenAI-shaped APIs.
実行環境: NVIDIA GPU workflow with built-in vLLM 0.26; the default 0.6B model is the lighter starting point and the 1.7B model trades memory for capability.
メモリ要件Measured startup: 4,620 MiB (0.6B) and 7,060 MiB (1.7B). The optional forced aligner adds roughly 3.3 GiB when first loaded.
速度 / RTFThe 0.6B model processed 1,320.936 seconds of 30-language audio in 32.376 seconds on an RTX 5070 Ti, or 40.80 times real time, with a 96.10% benchmark score.
You do not need to know which installer or GPU package you need. Answer two questions and we will give you the safest starting path from the official documentation.
123
1
Which operating system do you use?
2
Which graphics hardware do you have? “Not sure” is fine
03 / Your setup path
Safe first checkConfirm Docker can start a tiny test container.
docker run --rm hello-world
Docker says “Hello from Docker!”?Perfect. Your runtime works. Next, choose a lightweight project and follow its repository instructions.