WaveCap @WaveCapSD

I have Built a Mac app for heavy-duty video and audio transcription, and I need your brain to improve it. What matters most to you when transcribe massive video/audio files? 1. High precision transcription? 2. High speed transcription? 3. Speaker diarization? 4. AI summary? 5. Interactive AI chat to query the transcript? 6. Testing out multi AI models? 7. Integration of other tool/app? What’s missing or most critical for your workflow? Drop your thoughts below to help me shape the app! You may want to check out my app too apps.apple.com/app/wavecap-ai…

Why are you still paying monthly subscriptions or per-minute fees to transcribe your files? Most web services charge expensive rates just to run the exact same open-source speech models on their servers. If you are transcribing a 2-hour movie or bulk content, the cost adds up incredibly fast. WaveCap brings the power back to your local machine: Local Inference: Leverage your Mac's own unified memory and GPU to run premium STT models locally. 100% Private: No cloud servers, no API keys, and absolutely no data leaks—your files never leave your Mac. One-Time Lifetime Purchase: No monthly credits or recurring bills. Just pay once for unlimited transcription and free updates forever. Let's make professional AI transcription cheap, private, and accessible to everyone. apps.apple.com/app/wavecap-ai…

What actually is MLX, and why did we build WaveCap on top of it? 1. What is MLX? (Apple's NumPy) Developed by Apple, MLX is an open-source machine learning framework designed specifically for Apple Silicon - Unified Memory Integration: Standard ML frameworks require constantly moving data back and forth between system RAM and dedicated GPU memory. MLX is built around Apple Silicon's unified memory, meaning arrays live in a shared memory pool accessible by both CPU and GPU without memory-transfer overhead. - Familiar Developer Framework: It mirrors the APIs of NumPy for array manipulation and PyTorch/JAX for neural networks (mlx.nn), alongside automatic differentiation and gradient tracking. Beyond Python, it includes native Swift, C++, and C APIs. - Lazy Computation & Dynamic Graphs: Computations in MLX are lazy—it constructs a dynamic graph in the background and only materializes values when you explicitly read or print the result, allowing the engine to optimize operations behind the scenes. - Inference and Training in One Package: Unlike tools like Ollama or llama.cpp that primarily handle model inference, MLX supports full model training and fine-tuning (e.g., LoRA) on-device. Optimized for Open-Source LLMs: Through extensions like mlx-lm, it is used to download, quantize (compress to 4-bit or 8-bit), run, and fine-tune open-weights models (such as Llama, Qwen, and Mistral) directly on Mac hardware. 2. The Challenge of Local AI Running a complete, high-quality transcription pipeline locally on a computer is incredibly resource-demanding. If you stitch together separate, unoptimized libraries for signal processing, speech-to-text, and language models, your computer's RAM and CPU will quickly choke. 3. Why WaveCap Chose MLX MLX allows WaveCap to run a complete, modular local AI stack sequentially within a single, ultra-efficient memory footprint - Adaptive Vocal Isolation: If heavy noise is detected, we route raw audio through Demucs to isolate clean speech - Voice Activity Detection (VAD): We apply Silero VAD to slice audio into discrete speech intervals, stripping dead air to cap peak memory and eliminate hallucinations - Blistering-Fast Speech-to-Text: We run 20+ local models (like Whisper, Qwen3 ASR, and Parakeet) natively on the GPU - Local LLM Polishing & Translation: We use offline LLMs (like Gemma and Qwen) to remove filler words ("ums"), correct grammar, and translate transcripts into 47+ languages - Speaker Diarization: We run Sortformer to cluster and label persistent speakers throughout your media timeline The Verdict By leveraging MLX's Unified Memory, WaveCap executes this entire multi-phase pipeline sequentially with maximum speed and minimal system memory overhead No cloud dependencies, no API subscription fees, and zero data leaks. Just 100% private local AI running directly on your Mac. Try it out here: apps.apple.com/app/wavecap-ai… #MLX #UnifiedMemory #WaveCap #MacApp #AITranscription #NumPy #Swift #Qwen #Whisper #STT

@Guronnimo AI Transcription macApp powered by MLX apps.apple.com/app/wavecap-ai…

@lightsilver323 apps.apple.com/app/wavecap-ai…

MLX STT vs Whisper.cpp vs WhisperKit On-device Speech-to-Text on Apple Silicon is moving at warp speed, and developers are constantly asking us: "Why did WaveCap build on MLX instead of Whisper.cpp or WhisperKit?" Here is why MLX is the only framework that makes sense for the future of on-device AI transcription: 1. Breaking Free from Model Lock-in Whisper is legendary, but it is no longer the only game in town. - Whisper.cpp and WhisperKit restrict you strictly to OpenAI’s Whisper model family - By building on MLX, WaveCap can run 20+ different local speech-to-text architectures natively. We support Whisper, but we also run next-gen models like Parakeet, Qwen3 ASR, SenseVoice, and GLM-ASR 2. Unified Memory is a Superpower Apple Silicon relies on Unified Memory, and MLX is built from the ground up to exploit this architecture - Whisper.cpp and WhisperKit treat the CPU, GPU, and Neural Engine as separate silos - MLX allows us to run highly efficient, low-bit quantized models (like 4-bit Qwen3 ASR or GLM-ASR) directly on the GPU, maximizing speed while keeping our RAM footprint incredibly low 3. Powering a Complete Local AI Stack WaveCap is much more than a simple play button for transcription; it is a modular, multi-phase local AI pipeline We don't just transcribe. In one click, WaveCap can: - Isolate vocals and clean noise with Demucs - Detect voice activity via Silero VAD to strip out dead air - Transcribe your speech - Polish and translate raw text using local LLMs (Gemma & Qwen) - Cluster speakers with neural Sortformer Diarization MLX is the only framework that lets us run signal processing, STT, and LLMs under one unified, local memory architecture on your Mac 4. Absolute Future-Proofing In the open-source AI world, breakthrough models drop every week. - With CoreML or GGML-based engines, developers have to wait for custom binary conversion scripts or custom wrappers before they can run new models - With MLX, we can download and deploy quantized, state-of-the-art models directly from the Hugging Face Hub the second they are released The Verdict If you only need standard Whisper, other tools are great. But if you want a blistering-fast, modular, and future-proof app that unleashes the full stack of local AI—from vocal cleanup to LLM rewriting—MLX is the only way forward Experience the power of local MLX transcription today: apps.apple.com/app/wavecap-ai… #MLX #STT #SpeechToText #ASR #AutoSpeechRecognition #Whispercpp #WhisperKit #MacApp #WaveCap #LLM

Speed vs Accuracy Speed is great, but if your offline AI transcription is fast and wrong, did you actually save any time? Here is why WaveCap chooses elite accuracy over raw, unoptimized speed—and how it will save you hours of post-editing: 1. The "Editing Tax" Having used AI transcription for years, I noticed a painful pattern. Many local apps brag about processing files in seconds, but use low-accuracy models that hallucinate constantly. You end up spending more time editing and correcting the broken text than you would have spent just typing it manually. 2. Accuracy is the Real Speedometer At WaveCap, we prioritize accuracy first. But that doesn't mean we don't care about transcription speed. True speed is the total time from dragging your file into the app to having a perfectly polished, ready-to-use transcript. If you don't have to spend a single second fixing typos, that is real speed. 3. Breaking the Whisper Monopoly Look at the local transcription landscape: almost every app relies solely on OpenAI's Whisper models. But speech AI is moving incredibly fast. Over the last few years, state-of-the-art models have emerged that rival or beat Whisper in both accuracy and speed—like Qwen3 ASR and Parakeet. 4. Bring on the SOTA Models That’s exactly why we built WaveCap. We didn’t want to stick with Whisper alone. By leveraging Apple's native MLX framework, WaveCap gives you the freedom to choose from 20+ local speech-to-text models: - Whisper Large v3 & Large v3 Turbo - Qwen3 ASR (0.6B & 1.7B) - Parakeet TDT & CTC variants - Voxtral Realtime, GLM-ASR, & SenseVoice You can run smaller, lightning-fast models for quick drafts, or deploy heavy-duty, ultra-high-accuracy models for complex audio. Stop wasting your day fixing AI hallucinations. Run the world's best open-source speech models natively and privately on your Mac. Download WaveCap on the Mac App Store apps.apple.com/app/wavecap-ai… #AITranscription #WhisperAI #QwenASR #Qwen #Parakeet #Voxtral #MacApp #MLX

MacWhisper remains a gold standard, but if you want to experiment with advanced, ultra-modern STT models offline on Apple Silicon, I hope WaveCap gives you a fresh experience. 📥 Try it here: apps.apple.com/app/wavecap-ai… #IndieHackers (4/4)

As a Mac user, I’ve used MacWhisper for a long time—it’s truly a fantastic tool. But as new AI architectures dropped, I hit a few personal workflow bottlenecks with model variety, translations, and single exports. So, I decided to build an alternative: WaveCap. 👇 (1/4) #MacApp

WaveCap splits processing to maximize local efficiency: 1️⃣ Vocal isolation (Demucs) 2️⃣ VAD (Silero) to chunk audio & save memory 3️⃣ STT (Choose from 15+ models) 4️⃣ LLM text polishing & translation 5️⃣ Diarization (Sortformer) This unlocks true modular power. 👇(3/4) #OnDeviceAI

Whisper is great, but speech AI has evolved. I wanted to try state-of-the-art models like Qwen ASR, GLM, and Nemotron using Apple MLX. To do this, I had to completely re-architect the traditional workflow, breaking it down into modular steps for future-proofing. 👇(2/4) #MLX

WaveCap support the following AI Models STT: Whisper, Qwen3 ASR, GLM, Nemotron ASR, Voxtral Mini, Parakeet, Cohere, MOSS, SenseVoice Small Vocal Separation: Demucs, MDX Voice Detection: Silero VAD Diarization: Sortformer #WhisperAI #Qwen #GLM #Nemotron #Voxtral #Cohere #MOSS

🚀 Exciting news! I just launched WaveCap, a 100% private AI transcription app for macOS. 💻 Powered by Apple MLX with 50+ local models, it turns audio/video into text blazingly fast—completely offline. Try it out here: apps.apple.com/app/wavecap-ai… #MacApp #MLX #AI #Whisper #Qwen