เกี่ยวกับ StemSplit
ทำให้การแยกเสียงระดับมืออาชีพเข้าถึงได้สำหรับทุกคน
ภารกิจของเรา
เราเชื่อว่าทุกคนควรมีสิทธิ์เข้าถึงเครื่องมือเสียงระดับมืออาชีพ ไม่ว่าคุณจะเป็นโปรดิวเซอร์มือใหม่ คนรักคาราโอเกะ หรือครีเอเตอร์ StemSplit ให้ผลลัพธ์คุณภาพสตูดิโอโดยไม่ต้องมีงบประมาณสตูดิโอ เรากำลังพัฒนาอย่างเปิดเผย และคำแนะนำของคุณกำหนดแผนงานของเราโดยตรง
เราให้บริการใคร
- โปรดิวเซอร์เพลงและ DJ
- คนรักคาราโอเกะ
- ครีเอเตอร์และ YouTuber
- พอดแคสเตอร์และนักตัดต่อเสียง
- ครูสอนดนตรีและนักเรียน
- Developers building audio apps
Built on Serious Technology
StemSplit uses HTDemucs — a state-of-the-art hybrid transformer model developed by Meta Research and published at NeurIPS 2022. It's the same model researchers use, made accessible through a clean web interface.
HTDemucs Model
HTDemucs is a hybrid transformer/convolutional model that works in both the waveform and spectrogram domain simultaneously. This dual-domain approach is what gives it its accuracy advantage over older CNN-only models like Demucs v3 or Spleeter.
ONNX Export
We export the model to ONNX format — an open standard for machine learning interoperability. This means the model runs without PyTorch, making it lighter, faster, and portable. We've published the ONNX weights on HuggingFace for anyone to use.
Up to 6 Stems
Standard mode separates audio into vocals, drums, bass, and other. The 6-stem model additionally separates guitar and piano. Fine-tuned variants (htdemucs_ft) are available for single-stem extraction with higher accuracy.
No GPU Required
Processing happens entirely in the cloud. You upload a file, we run inference server-side, and you download the separated stems. No software to install, no GPU needed on your end.
Who Built This
StemSplit was built by a solo developer who kept running into the same frustration: existing vocal removal tools were either buried behind monthly subscriptions you'd forget to cancel, or they used credits that expired before you needed them again.
The goal was simple — build the tool that should already exist. One where you pay a fair price for what you actually use, your balance never expires, and the output quality is good enough that you don't need to try three different services.
It's an indie product. There's no VC money, no growth team, no sales calls. Just a tool that works, priced honestly, maintained by one person who uses it too. The roadmap is driven by users — if you have a suggestion, it genuinely gets read.
Built in public
Development happens openly. Packages are published on GitHub, PyPI, npm, and HuggingFace. The benchmark methodology is public. When things break, they get fixed and the fix is visible.
More Than a Website
StemSplit is also a developer platform. The same separation engine that powers the web app is available through multiple distribution channels.
REST API
Full programmatic access to stem separation. Submit jobs, poll status, download results. Used by developers building their own tools on top of StemSplit.
Python Package
stemsplit-python on PyPI. Run separation locally or call the API from Python scripts and data pipelines.
CLI Tool
stemsplit via Homebrew. Separate audio files from the terminal with a single command. Useful for batch processing and shell scripts.
n8n Node
n8n-nodes-stemsplit on npm. Drag-and-drop stem separation inside n8n automation workflows — no code required.
MCP Server
stemsplit-mcp on npm. Expose stem separation as a tool that AI agents (Claude, GPT, etc.) can call directly via the Model Context Protocol.
Zapier Integration
Connect StemSplit to thousands of apps via Zapier. Trigger separations from Google Drive uploads, new emails, form submissions, and more.
Open Methodology
We built and published a benchmark dataset on HuggingFace that evaluates multiple HTDemucs model variants across real music tracks. The evaluation uses Signal-to-Distortion Ratio (SDR) — the standard academic metric for source separation quality — alongside listening tests across different musical genres and stem types.
View the benchmark dataset on HuggingFaceEvaluation metric
What is SDR?
SDR (Signal-to-Distortion Ratio) measures how cleanly a model separates a target source from the mix, in decibels. Higher is better. It's the same metric used in the SiSEC Music Separation benchmark, the standard academic evaluation for this problem.
Typical SDR scores (HTDemucs)
Source: StemSplit benchmark dataset on HuggingFace. Higher SDR = cleaner separation.
Why We Charge the Way We Do
Most audio tools use subscriptions because subscriptions are good for the business — they generate predictable revenue even from users who barely log in. That's not how we want to operate.
StemSplit uses a pay-as-you-go credit model. You buy minutes of processing time and they never expire. If you separate one song a month, you pay for one song. If you batch-process an album, you pay for that. There's no plan to upgrade to, no features gated behind a higher tier.
The free tier gives every new user 5 minutes to try the product for real — not a watermarked preview or a 30-second clip. Five full minutes of actual output.
How it works
- Credits never expire — ever
- No subscription required
- 5 free minutes on signup, no card needed
- Same quality at every price point
- Full API access included
สิ่งที่เราเชื่อ
ความเรียบง่ายมาก่อน
เทคโนโลยีซับซ้อนควรใช้งานง่าย อัปโหลด ประมวลผล ดาวน์โหลด — แค่นั้น
ราคาโปร่งใส
ไม่ต้องสมัครสมาชิก เครดิตไม่หมดอายุ จ่ายเฉพาะที่คุณใช้ เมื่อคุณใช้
พัฒนาอย่างเปิดเผย
เรากำลังพัฒนาอย่างเปิดเผยร่วมกับชุมชนของเรา คำแนะนำและความคิดเห็นของคุณไม่เพียงถูกรับฟัง — มันคือแผนงานของเรา