Whisper cpp gpu

Whisper cpp gpu. dk/ 👉 Subtitle Edit on GitHub whisper. In this article, you will find the following: Whisper benchmarks and results opencl: enable OpenCL support. Model: ggml-large-v3, lower is better. 120+xpu and both model loading time and decoding have a "good" speed now. Aug 1, 2023 · Now build whisper. Oct 19, 2022 · Random (slightly adjusted) ChatGPT (GPT v4) advice that helped me. We're excited to announce WhisperScript v1. Recently, Georgi Gerganov released a C++ port optimized for CPU and especially Apple Silicon Platform. However, as the GPUs inference speed is so much faster than real-time anyways (around 0. To get whisper up and runnig with cuda I had to do the following steps. cpp running on a MacBook Pro M1 (CPU only) Hope you find this project interesting and let me know if you have any questions about the implementation. BLAS CPU support via OpenBLAS Collaborate on models, datasets and Spaces. Author. Try to run it on that Veritasium video and see it for yourself. generate , including num_beams for beam-search, return_timestamps for segment-level timestamps, and prompt_ids for Nov 2, 2022 · Hello there, In principle you should be able to apply TensorRT to the model and get a similar increase in performance for GPU deployment. Copy paste your audio file (s) that you want to convert into the data folder. cpp on a Jetson Nano for a real-time speech recognition task. Apr 16, 2023 · is there a way to run whisper. Sep 21, 2022 · small 244 M small. CPP is always much faster than Whisper on CPU, over 6 times faster for the tiny model up to over 7 times faster for the large one. Jan 8, 2024 · Hi, I am a nixOS beginner for about a month. 👉 Official Subtitle Edit Website 👉 https://nikse. だいたい GPU の 10 倍くらい時間 Upstream whisper. cpp's log output and sending it to the log backend. Mar 6, 2012 · In this Subtitle Edit Tutorial, we'll look at the different Whisper Modes available in Subtitle Edit. Clone the Repository. This implementation is up to 4 times faster than openai/whisper for the same accuracy while using less memory. 3× slower). bin" model weights. faster-whisper is a reimplementation of OpenAI's Whisper model using CTranslate2, which is a fast inference engine for Transformer models. cpp directory, run: We would like to show you a description here but the site won’t allow us. cpp : WASM example. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info. 進入whisper環境 Apr 20, 2023 · I will share benchmarks and test results. For example, Whisper. Link al articulo mi blog:https://construyendoachisp Port of OpenAI's Whisper model in C/C++. Plus, code to transcribe YouTube videos using whisper. To be complete it was not working for me without others steps. _ext. cpp 项目是将 Whisper 移植到 C/C++ 中，而今天介绍的 Const-me/Whisper 项目则是 whisper. Mar 21, 2024 · Whisper. cpp development by creating an account on GitHub. swiftui: SwiftUI iOS / macOS application using whisper. I can run the stream method with the tiny model, but the latency is too high. The ggml library is one of the first library for local LLM interference. We're using an Nvidia GPU with CUDA support, so our Supports transformers, GPTQ, AWQ, EXL2, llama. まずは openai-whisper. cpp repository: Abstract: Whisper is one of the recent state-of-the-art multilingual speech recognition and translation models, however, it is not designed for real-time transcription. cpp (CPU) and openai-whisper (GPU), as well as code to visualize the benchmark results with matplotlib. It runs slightly slower than Whisper on GPU for the small, medium and large models (1. Some updates: Dec 12, 2023 · You need to run WHISPER_CUBLAS=1 make stream. for those who have never used python code/apps before and do not have the prerequisite software already installed. Dec 8, 2022 · With much less RAM and slower M1 (non-Pro) processors, you can see the numbers declining a lot. Requires calling. bobqianic added the question Further information is requested label Dec 15, 2023. The version of Whisper. cpp Distil-Whisper can be run with the Whisper. …. md files in Whisper. wav" --model medium --device cuda. set_per_process_memory_fraction(), so it may not be actually reflecting what happens in e. g. net 1. Read README. cpp does not treat OpenCL as a GPU, so it is always enabled at runtime. with model file in page cache. a GPU with exactly 5 GB VRAM. It has no dependencies, low memory usage, excellent performance, supports multiple technologies and platforms, supports mixed precision and integer quantization and other advantages. huapingchen mentioned this issue on May 13, 2023. 2008. CPP is faster than Whisper on GPU for the tiny (1. / ``` Run all the examples as usual. medium 769 M medium. BLAS CPU support via OpenBLAS. モデルによっては，git-lftが必要になるかもしれません．. Sep 23, 2022 · daleh October 29, 2022, 8:00pm 16. anaconda安装无脑下一步就好 Jan 31, 2023 · この Whisper. The English-only models were trained on the task of speech recognition. 👍 1. Here are the steps for creating and using a quantized model: Port of OpenAI's Whisper model in C/C++. The JAX code is compatible on CPU, GPU and TPU, and can be run standalone (see Pipeline Usage ) or as an inference endpoint (see Creating an Endpoint ). 59 ms. load_model("base", device="cuda") # If you are loading Whisper using Port of OpenAI's Whisper model in C/C++. Minimal example running fully in the browser. 1 is based on Whisper. 10. 5 seconds for 30 seconds of real-time audio), this would only be useful if you was transcribing a large amount of audio (podcasts, movies, large amounts of audio files Apr 12, 2024 · Apr 12, 2024. cpp that referenced this issue on Nov 19, 2023. Oct 19, 2022 · I tried Whisper on a Jetson tx2 development kit (ubuntu 18. 2. まずは本家 openai-whisper 使います. I was running the desktop version of Whisper using the CMD prompt interface successfully for a few days using the 4GB NVIDIA graphics card that came with my Dell, so I sprang for an AMD Radeon RX 6700 XT and had it installed yesterday, only to discover that Whisper didn't recognize it and was reverting my my CPU. whisper. . Whisper is a Transformer based encoder-decoder model, also referred to as a sequence-to-sequence model. モデルの用意. Share Add a Comment Mar 4, 2023 · 最終的に, fp16 だと GPU メモリ 4. cpp repo to download one of the models. In a provisional benchmark on Mac M1, distil-large-v3 is over 5x faster than Whisper large-v3, while performing to within 0. tamo mentioned this issue on Nov 19, 2023. en medium ~5 GB ~2x. Dockerfile has some issues. flac. For more control over the generation parameters, use the model + processor API directly: Ad-hoc generation arguments can be passed to model. This may have performance implications. Install whisper Mar 30, 2023 · faster-whisper とは. The models were trained on either English-only data or multilingual data. metal: enable Metal support. This update adds a bunch of improvements to the visualization, playback, editing, and exporting of your transcripts. cpp and server of llama. 3. Whisper. cpp cmake -B build -DWHISPER_CLBLAST=ON cmake --build build -j --config Release Run all the examples as usual. is_available() is False. The latter is not absolutely necessary but added as a workaround because the decoding logic assumes the outputs are in the same device as the encoder. android: Android mobile application using whisper. Apr 24, 2024 · Developers can now integrate ChatGPT and Whisper models into their apps and products through our API. cpp on gpu? Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community. Whisper Sample Code Apr 9, 2024 · I ran into the same problem. . GPTQ-for-LLaMa - 4 bits quantization of LLaMA using GPTQ ggml - Tensor library for machine learning alpaca. 「音声認識モデル Whisper の推論をほぼ倍速に高速化した話」を参考に fp16 化 + no_grad Oct 7, 2022 · Whisper CPP がありました! libtorch や ONNX など使わず, 自前の ggml で model の forward 計算と, 自前で C++ で FFT 処理 (MEL spectrogram 計算) + Token decode していてしゅごい. Drop-in replacement for OpenAI running on consumer-grade hardware. time whisper --language en --model tiny. The English-only models were trained on the task of speech Now build whisper. Begin by cloning the dedicated Whisper. It provides high-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model running on your local machine. Each version of Whisper. We would like to show you a description here but the site won’t allow us. 開啟Anaconda Prompt. cpp 项目，以及该项目的 GUI 版本 ：whisperDesktop 。 May 5, 2023 · windows本地搭建openai whisper并开启NVIDIA GPU加速 需要的工具. en small ~2 GB ~6x. cpp make clean WHISPER_CLBLAST=1 make -j CMake: cd whisper. jacob-salassi mentioned this issue on May 14, 2023. bobqianic added the solution This issue contains a potential We would like to show you a description here but the site won’t allow us. cpp is to enable LLM inference with minimal setup and state-of-the-art performance on a wide variety of hardware - locally and in the cloud. en. ← WavLM XLS-R →. 👍 2. sh: Livestream audio Mar 9, 2023 · Hey, thanks for sharing this here! I'm the author of faster-whisper and CTranslate2. in the terminal, and create a new folder called data. Oct 12, 2023 · I am trying to run whisper. Sep 22, 2022 · After doing this I was able to run the following command ( choose whatever model fits your needs best ): whisper "audio. sh: Helper script to easily generate a karaoke video of raw audio capture: livestream. Step 2 : 安裝whisper步驟. 3 xtmu, gparvind, and LaptopDev reacted with thumbs up emoji. Next, you need to fetch a Whisper model converted into . Plain C/C++ implementation without any dependencies. cpp into your transcription workflow effectively, here’s a stepwise guide: 1. Requires calling Feb 7, 2024 · Jianningyuan commented on Feb 7. cpp is compiled and ready to use. To achieve good performance, you need an Nvidia CUDA GPU with > 8 GB VRAM. この実装は、openai / whisperよりも最大4倍高速で、同じ精度で、より少ないメモリを使用します。. 3× faster). cpp repository to obtain the source code. cpp, and there is roughly a 1. 輸入conda create — name whisper python=3. @ggerganov, is there a reason why it's not included in the code? I'm completely unaware of the technical details of this topic, but I'd like to use GPU by We also provide an end-to-end Google Colab that benchmarks Whisper against Distil-Whisper. Want to try it yourself? Use my code to run your own tests. pythonの Jul 13, 2023 · doesn't work with arch=any OR arch=-arch=compute_61 for my GTX 1080 in Makefile make clean WHISPER_CUBLAS=1 make stream -j I whisper. wav -t 8. Nov 17, 2023 · Yes. Requirements: Full admin rights on your computer. You need to use the 16 kHz sampling rate wav file. A PC with a CUDA-capable dedicated GPU with at least 4GB of VRAM (but more VRAM is May 29, 2023 · AI字幕神器whisper最全中文攻略. Powered by OpenAI's Whisper. bin -f rhasspy/setkitchenlights20percent. Category. cpp/models にあるREADMEにhuggingfaceのモデルを使用する場合の流れが書いてあるので，それに従います．. ggml format. Transcribe an audio file, alternatively specifying language, model, and device. RTX 3090 Advantage. cpp package with the original sequential long-form transcription algorithm. 11), it works great with CPU. 5B params). Buzz is better on the App Store. This is more of a real world test with actual work loads to be handled. I use a modified nix file, and add cudaPackages libcublas cudatoolkit in buildInputs and cuda_nvcc in nativeBuildInputs, also add env = { WHISPER_CUBLAS = "1"; }. The VRAM requirements are from simulations using torch. faster-whisperは、トランスフォーマーモデルの高速推論エンジンであるCTranslate2を使用したOpenAIのWhisperモデルの再実装です。. Feb 2, 2024 · Whisper: The original Whisper model is implemented in Python and supports running on both the CPU and the GPU. Runs gguf, trans Whisper is a Transformer based encoder-decoder model, also referred to as a sequence-to-sequence model. cpp with CLBlast support: ```Makefile:cd whisper. cpp with CLBlast support: Makefile: cd whisper. 5. o and also -lcudart -lcublas. 1. 500. whisper是OpenAI公司出品的AI字幕神器，是目前最好的语音生成字幕工具之一，开源且支持本地部署，支持多种语言识别（英语识别准确率非常惊艳）。. 7. 8, PyTorch 2. I tested openai-whisper-cpp 1. Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks. To integrate Whisper. Quantized models require less memory and disk space and depending on the hardware can be processed more efficiently. so). Nov 16, 2022 · The code above uses register_forward_pre_hook to move the decoder's input to the second GPU ("cuda:1") and register_forward_hook to put the results back to the first GPU ("cuda:0"). cpp build info: I UNAME_S: Linux I UNAME_P: x86_64 I UNAME Skip to content Whisper. share/whisper-cppのディレクトリをexportすればOK ※brewのformulaを読むとggml-metal. In this situation, as only exception to the description above, the value specified for '--gpu-architecture' may be a 'real' architecture (such as a sm_50), in which case nvcc uses the specified 'real' architecture and May 11, 2023 · En este vídeo veremos como instalar y usar whisper. cpp 在 Windows 上的实现，并增加了显卡的支持，使得速度大幅提升。 Apr 6, 2024 · opencl: enable OpenCL support. As of now (2023/05/01), here’s an example flow to do this: Clone the whisper. Might help others as well, YMMV:""" To resolve this issue, you should modify the instantiation of the ctranslate2. cpp: whisper. Mar 27, 2024 · Then we simply call the Whisper model using our GPU instead of CPU: # If you are loading Whisper using GPU gpu_model = whisper. CPUとGPUの両方で8 Jan 27, 2024 · whisper. For example, I applied dynamic quantization to the OpenAI Whisper model (speech recognition) across a range of model sizes (ranging from tiny which had 39M params to large which had 1. 今回は以下のモデルを使用したいと思います．. this will modify the gcc command so that it includes ggml-cuda. $ . 0 is based on Whisper. on Apr 19, 2023. pip install -U openai-whisper. cpp. cpp en tu propio ordenador para pasar de audio a texto. ) I am pretty sure that the box I am connecting to via ssh has 20 CPU cores and 2 GPU cores however the GPUs are not recognized by Jax. 0, CUDA 10. This is the smallest and fastest version of whisper model, but it has worse quality comparing to other models. f589894. Mar 13, 2024 · Table 1: Whisper models, parameter sizes, and languages available. If you have other types of files. whisper开源 Apr 24, 2023 · 这里提供 Whisper 及 Whisper. segmentation fault when use Core ML #919. 5 GB くらいあれば動きます. Apr 22, 2023 · Step 1 : 建議使用Anaconda安裝，請於下圖下載Ananconda. cpp ; mkdir build ; cd buildcmake -DWHISPER_CLBLAST=ON . mp3 -ar 16000 -ac 1 -c:a pcm_s16le output. cpp folder in Finder using open . I tried the CuBLAS instructions, but I could not get it to work (maybe my bad or GPU incompatibility) I would appreciate it if you guys could give me a tip or some advice. en --device xpu tests/jfk. Encoder processing can be accelerated on the CPU via OpenBLAS. It's also possible for Core ML to run different portions of the model in different devices Jan 5, 2023 · With ffmpeg installed, you can now open your whisper. The existing CPU-only implementation achieves this goal - it is bloat-free and very simple. leuc on May 15, 2023. RTX 3090. to get started. cpp repo; From the whisper. The efficiency can be further improved with 8-bit quantization on both CPU and GPU. 2. Open your terminal again in the whisper. OpenAI released Whisper in September 2022 as Open Source. please use ffmpeg -i input. nvim: Speech-to-text plugin for Neovim: generate-karaoke. Thanks! Jan 8, 2023 · 結論から言うと，whisper. Jul 14, 2023 · I can confirm that the issue of not running in GPU on macOS M1 is still present. When compiling stuff with CUDA support you need to distinguish between the compile phase and the runtime phase: When you build the image with docker build without mapping a graphics card into the container the build should link against the CUDA library stubs (e. Switch between documentation themes. 8× faster) and base models (1. 8% WER over long-form audio. cpp is a lightweight intelligent speech recognition library, which is a port of the OpenAI Whisper model. Just re-tried with v1. 3, python 3. Dec 20, 2022 · For CPU inference, model quantization is a very easy to apply method with great average speedups which is already built-in to PyTorch. It does indeed: that runs slower than CPU alone. dll files to the release, for added convenience. bobqianic added the question label on Nov 18, 2023. GPUを使いたい場合は、 . May 11, 2023 · Download OpenAI's Whisper. The large-v3 model is the one used in this article (source: openai/whisper-large-v3). The reason why it went to “Welsh” on your accent is that it does autodetect language and output in that language. Nov 10, 2022 · Cuda is apparently unsupported on M1 macs, and if I try --device mps I get a warning that The operator 'aten::repeat_interleave. Apr 7, 2024 · セットアップ. Get a Mac-native version of Buzz with a cleaner look, audio playback, drag-and-drop import, transcript editing, search, and much more. Implicitly enables hidden GPU flag at runtime. 6. After a good bit of research I found that the main-cuda. load_model("small", device="cuda") 发现仍然无法运行，提示我 RuntimeError: Attempting to deserialize object on a CUDA device but torch. cpp and llama. cpp folder, and run the following command: mkdir -p output; Apr 1, 2023 · In this video, I'll show you How to Use Whisper via GPU in Subtitle Edit. 000 --> 00:07. cpp run the live stream using following code: (GPU, step set to 700 works fine) I get "hallucinations" when not talking :robot: The free, Open Source OpenAI alternative. No GPU required. However, the patch version is not tied to Whisper. That is Whisper Open AI vs Whisper CPP vs Whisper Const May 10, 2024 · iOS mobile application using whisper. It’s a pure C library that converts models to run on several devices, including desktops, laptops, and even mobile device - and therefore, it can also be considered as a tinkering tool, trying new optimizations, that will then be incorporated into other downstream projects. Complie Whisper. Nov 6, 2022 · Will ggml / whisper. Feb 1, 2023 · Whisper. Nov 22, 2022 · model = whisper. 04, Nvidia Jetpack 4. 5x to 2x speedup compared to Openai’s Whisper model, but on my Intel i7-6700T only the tiny model sort of usable. cpp: Whisper. 2) with a 2 sec mp3 audio and it took 12 sec to transcribe with --language en --model tiny parameters. Not Found. cpp, 19 minutes audio transcribe, with Chinese Mandarin and English spoken. tamo added a commit to tamo/whisper. The availability of advanced technology and tools, in particular, AI is increasing at an ever-rapid rate, I am going to see just how easy it is to create an AI-powered real-time The main goal of llama. openblas: enable OpenBLAS support. Flash + language support (ref ggerganov#2) …. cpp 项目的模型文件下载及简单使用，后附 Whisper. Load Time. Install PaddleSpeech. This major release includes the following changes: Full GPU processing of the Encoder and the Decoder with CUDA and Metal is now supported Apr 16, 2022 · WARNING - No GPU/TPU found, falling back to CPU. Apr 2, 2023 · whisper. self_int' is not currently supported on the MPS backend and will fall back to run on the CPU. We’re on a journey to advance and democratize artificial intelligence through open source and open science. 0 and Whisper A Step-by-Step Guide to Whisper. libcuda. net is the same as the version of Whisper it is based on. Contribute to ggerganov/whisper. large 1550 M N/A large ~10 GB 1x. May 2, 2023 · ちなみにwhisper高速化兄弟のwhisper-jaxだとGPUをほぼに使ってくれるので、RTX3600環境で10分の音源を3分程度で書き起こししてくれました。 GPU環境があるのであれば、whisper-jaxが最速の可能性があります。 faster-whisperもあるので、速度比較してみたいです。 May 20, 2023 · whisper. cpp - Locally run an Instruction-Tuned Chat-Style LLM faster-whisper - Faster Whisper transcription with CTranslate2 Jun 21, 2023 · Purpose: These instructions cover the steps not explicitly set out on the main Whisper page, e. In this paper, we build on top of Whisper and create Whisper-Streaming, an implementation of real-time speech transcription and translation of Whisper-like models. wav) Click on the "Transcribe" button to start the transcription. Nov 15, 2023 · Three weeks ago I discovered whisper. Also, I've tried the solution from @diaojunxian and it works! Inference is much faster when it's run on GPU. Assuming you want to use your GPU, the device should likely be 'cuda', and you may need to adjust the compute_type and othe Mar 28, 2023 · in theory if you are succeed doing the Core ML models you can have full advantage of any number of CPU, GPU and RAM allocated on your device because Core ML supports all the compute units available in your device: CPU, GPU and Apple's Neural Engine (NE). make cleanmake -jcp bin/* . cpp 的模型转换工具及一些其他必要组件等。 Whisper模型下载及使用Whisper的安装方法： 命令行安装，可以使用 pip 直接安装、更新：（如果友友看不明白pip命令那么直接跳到Whi Mar 4, 2023 · Author. I was a bit surprised by the benchmark results on this CPU. cuda. It is implemented in C/C++ and runs only on the CPU. Oct 1, 2022 · Port of OpenAI's Whisper model in C/C++. Follow the steps in the whisper. anaconda：python环境管理工具 chocolatey：windows包管理工具. cpp を知った際は、うちの初代ラズパイBでも動作するかなと期待したのですが、それなりにメモリは食うようなので断念。 ただ、GPU がない環境でも動くことが確認できたので、これからも色々試してみたいと思います。 Nov 19, 2023 · Whisper. cpp support CUDA / GPU? One of the main goals of this implementation is to be very minimalistic and be able to run it on a large spectrum of hardware. I was not able to use CUDA due to compatibility issues with pytorch 2. [00:00. cpp supports integer quantization of the Whisper ggml models. Self-hosted, community-driven and local-first. Also, when running whisper, my GPU hovers around 40-50% utilization, while running whisperX pushes it up to >95% utilization. b6344df. CMake:cd whisper. net is tied to a specific version of Whisper. I’ve been testing Whisper. It is suitable for scenarios that require real Compared to OpenAI's PyTorch code, Whisper JAX runs over 70x faster, making it the fastest Whisper implementation available. May 4, 2023 · If no value for option '--gpu-code' is specified, then the value of this option defaults to the value of '--gpu-architecture'. Faster examples with accelerated inference. Steps for getting started: Clone the Whisper. Make sure that the server of Whisper. It was trained on 680k hours of labelled speech data annotated using large-scale weak supervision. cpp using make. Closed. I want to test it with CUDA GPU for speed. anandijain pushed a commit to anandijain/whisper. Whisper object in your code to correctly match the expected constructor arguments. CPU 推論はいろいろ CPU 専用命令使ってもそれなりにはかかります. 1, an update to our Electron desktop Whisper implementation that introduces a lot of new features to speed up your transcription workflow. cpp that referenced this issue on Apr 27, 2023. cpp (GGUF), Llama models. This repository comes with "ggml-tiny. Dec 7, 2022 · OpenAIの高性能な音声認識モデルであるWhisperを、オフラインでかつGPUが無くても簡単に試せるようにしてくれたリポジトリを知ったのでご紹介。 Apr 11, 2023 · Windows11でPython版のWhisperを使いたかったけどPythonに触るのも久しぶりだったので色々調べながら。備忘録として残しておきます。 Jul 26, 2023 · 前一篇文章中 曾经介绍过 OpenAI 的开源模型 whisper，可以执行 99 种语言的语音识别和文字转写。 但是 whisper 模型占用计算资源多，命令行使用门槛高，所以还介绍了 whisper 模型的 c++ 实现：whisper. Thanks, this helped me too. M1 Max 24c GPU. 940] And so my fellow Americans ask not what your country can do for you. Upstream whisper. So you can let it do this, and it should work, if not, then prompt it in that language. cpp and tried it on my 6 year old desktop, CPU: AMD Ryzen 5 1600 Six-Core Processor, GPU: Radeon RX 460/560D, Video memory: 2048MB, gfx803, running Arch Linux based Manjaro. brew install whisper-cpp. 2 (from unstable and 23. 这篇文章应该是网上目前关于Windows系统部署whisper最全面的中文攻略。. /main -m models/ggml-tiny. @acoloss Try the prompting techniques from myself and @RonaldGRuckus, this should fix the “Welsh” problem. metalをコピーしていたのでこれを指定している。 Closed. 0. Usage instructions: Load a ggml model file (you can obtain one from here, recommended: tiny or base) Select audio file to transcribe or record audio from the microphone (sample: jfk. This is Unity3d bindings for the whisper. wav. cpp 1. Overview. cppmake cleanWHISPER_CLBLAST=1 make -j. cpp is a custom inference implementation of the Whisper model. ggerganov#1517 Redistribute CUDA DLLs. I a, able to run my raspberry pi 4 whisper. 13. Acquire the Whisper Model. We should add these . Whisper and whisperX also splits it up internally, but has mechanism to fix the boundaries and so are much better. whisper-cpp-log: allows hooking into whisper. 但我的设备是有GPU的，tensorflow可以正常调用GPU。 Mar 12, 2023 · Whisper 是一个由 OpenAI 训练并开源的神经网络，在英语语音识别方面的稳健性和准确性接近人类水平。 whisper. cv sa wh ys su ln vp wx wd zq