Soundtrap uses machine learning (sometimes referred to as AI) in a small number of optional tools for creating music and finding sounds.
- No chatbots or conversational AI (except support tickets submitted by teachers and administrators)
- No automated grading, student profiling, or decision-making about students
- No use of student or school data to train AI models
Machine learning features
| Feature | What it does | Available to | Model, where it runs |
|---|---|---|---|
| Transcription | Speech to editable text | Students, teachers | Google Speech-to-Text (Google's service) |
| Noise reduction | Removes background noise | Students, teachers | RNNoise, run by Soundtrap |
| Generate Drums | Creates a drum pattern | Students, teachers | Google Magenta, in the browser |
| Key detection | Finds a recording's key | Students, teachers | Spotify Basic Pitch, in the browser |
| Stem Splitter | Splits a song into vocals, drums, bass, other | Teachers only | Meta Demucs, run by Soundtrap |
| Top Picks | Suggests loops that fit the loops already in the project | Students, teachers | Soundtrap's own model, run by Soundtrap |
| Loop search | Finds loops that match what the user types | Students, teachers | Google Gemini (Google's service) |
All Soundtrap services are hosted on Google Cloud. "Google's service" means the data is processed by Google's own AI model; "run by Soundtrap" means only Soundtrap's code processes it.
Output is saved in the user's project like any other content. Users can edit or delete it at
any time.
Safeguards
- Only on request. Machine learning runs only when a user uses a tool, such as opening the loop browser or searching for loops. Nothing processes student audio automatically.
- Limited output. No tool generates open-ended content. Search text is only used to find matching loops in Soundtrap's catalog. Each tool returns a specific result: text from the user's own speech, cleaned or separated audio, drum notes, a musical key, or a list of loops.
- Profanity filter. Transcripts have profanity masked for English and most other supported languages.
- User control. All output is saved in the user's project and can be edited or deleted. Transcripts are fully editable to correct mistakes.
- Teacher oversight. On request, schools can enable teachers to review student projects, including any machine-learning output.
- Role restrictions. The Stem Splitter is available to teachers only.
- Data protection.
- No student data is used to train models. Top Picks is trained only on how non-school users combine loops.
- Transcription audio is processed by Google Cloud with data logging turned off, and no account details are sent.
- Loop search text is processed by Google Cloud to find matching loops, and no account details are sent. All other tools run on Soundtrap's servers or in the browser.
- Temporary processing files are deleted automatically within 7 days.
Comments
0 comments
Article is closed for comments.