Flash News: Official Desktop Version of DeepSeek Harness is Here!
I took it for a test drive right away, and two major highlights stood out:
- Seamless Migration: If you previously used the command-line interface (CLI) version of DeepSeek Harness, all your original configuration files and settings are automatically inherited. Zero re-setup required!
- Built-in Voice Input: The team thoughtfully introduced a new "Voice Input" feature (plugin), which makes daily interactions noticeably smoother and more efficient.
DeepSeek Harness is rapidly evolving and getting better with every update!
公式サイト:https://www.deepseek.com/harness/


Unexpected Discovery: Unearthing a Hardware Hidden Gem
While setting up the Voice Input plugin, I noticed it utilizes a model named SenseVoice under the hood. Driven by curiosity, I dug deeper into the hardware ecosystem supporting it and stumbled upon a intriguing piece of gear: the "M5Stack AI-8850 LLM Acceleration M.2 Kit 8GB (AX8850)".
This M.2 card provides an all-in-one local deployment stack for various AI models, including the full suite of AI audio models I'm particularly interested in.
Here is a quick breakdown of the 5 core AI voice models featured in this setup:
1. Whisper
- Developer: OpenAI
- Core Function: Automatic Speech Recognition (ASR / Speech-to-Text).
- Highlights: An end-to-end multilingual speech recognition model built on the Transformer architecture. It features exceptional noise robustness and zero-shot learning capabilities across 99+ languages, making it ideal for meeting transcriptions and video subtitle generation.
2. MeloTTS
- Developer: MIT & MyShell.ai
- Core Function: Text-to-Speech (TTS).
- Highlights: A lightweight, high-quality multilingual speech synthesis library. Its primary advantage is extremely low compute overhead—enabling real-time voice synthesis even on standard CPUs without a discrete GPU. It's a perfect fit for edge computing and low-cost deployments.
3. SenseVoice
- Developer: Alibaba Tongyi Lab (FunAudioLLM Team)
- Core Function: Multilingual Audio Understanding Foundation Model.
- Highlights: Beyond high-accuracy ASR, it supports Speech Emotion Recognition (SER) and Audio Event Detection (AED—e.g., applause, laughter, coughing). Its "Small" variant utilizes a non-autoregressive architecture that processes 10 seconds of audio in just ~70ms (about 15x faster than Whisper-Large), making it ideal for real-time interactive apps.
4. CosyVoice2
- Developer: Alibaba Tongyi Lab
- Core Function: High-Fidelity Speech Synthesis & Voice Cloning.
- Highlights: A next-generation speech generation LLM. It supports ultra-low latency (150ms) bidirectional streaming synthesis and zero-shot voice cloning using just a few seconds of reference audio. Users can also control emotion and prosody using natural language prompts, delivering remarkably human-like audio quality.
5. 3D-Speaker-MT
- Developer: Alibaba Tongyi Lab
- Core Function: Speaker Verification & Diarization.
- Highlights: An open-source multimodal speaker recognition toolkit. Beyond identifying *who* is speaking, it performs precise speaker diarization to segment and label different speakers in multi-party conversations. It also handles overlapping speech detection and integrates visual cues (like lip movement) for accurate multi-modal tracking in smart meeting systems.
Summary
In short, the core building blocks for an advanced AI voice system are neatly structured across three key pillars:
- Whisper & SenseVoice handle "Listening & Comprehension" (accurate transcription and emotion/context awareness);
- MeloTTS & CosyVoice2 handle "Expression" (lightweight execution and ultra-realistic speech synthesis/cloning);
- 3D-Speaker handles "Identification" (distinguishing and isolating different speakers).
With hardware and software integrating so seamlessly, the local AI audio ecosystem is maturing fast—paving the way for next-level interactive experiences on both desktop and edge devices!

