Speech Data Annotation Task
Budget / Salary₹12,500–37,500
TypeFreelance project
LocationRemote
Posted57 minutes ago
I need to turn a batch of existing audio files into fully-labeled training data for an automatic speech-recognition project. For every recording you will:
• note the language used and, when possible, the regional accent
• transcribe the spoken words verbatim
• tag the prevailing emotion (neutral, happy, angry, etc.)
• rate background-noise level and overall recording quality
Speaker-turn timing is not required; the focus is strictly on the language, accent, words, emotion, and environmental quality of each file.
You will receive the audio in WAV and MP3 formats together with a template (CSV or JSON—choose the format you prefer) that lists one row per clip. Simply fill the template with your labels and return it along with the completed transcripts. All annotations must line up to the millisecond offsets already provided in the template.
Acceptance criteria
• 98 % transcription accuracy when checked against spot audits
• Consistent use of the emotion and accent taxonomies provided
• Noise and quality ratings that match reference examples within ±1 category
Feel free to work in the tool you know best—Praat, ELAN, Audacity, or any browser-based annotator—as long as the output conforms to the template. A short pilot batch will be reviewed before green-lighting the full run so we can align on style and tagging decisions early.
• note the language used and, when possible, the regional accent
• transcribe the spoken words verbatim
• tag the prevailing emotion (neutral, happy, angry, etc.)
• rate background-noise level and overall recording quality
Speaker-turn timing is not required; the focus is strictly on the language, accent, words, emotion, and environmental quality of each file.
You will receive the audio in WAV and MP3 formats together with a template (CSV or JSON—choose the format you prefer) that lists one row per clip. Simply fill the template with your labels and return it along with the completed transcripts. All annotations must line up to the millisecond offsets already provided in the template.
Acceptance criteria
• 98 % transcription accuracy when checked against spot audits
• Consistent use of the emotion and accent taxonomies provided
• Noise and quality ratings that match reference examples within ±1 category
Feel free to work in the tool you know best—Praat, ELAN, Audacity, or any browser-based annotator—as long as the output conforms to the template. A short pilot batch will be reviewed before green-lighting the full run so we can align on style and tagging decisions early.
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.