Posted 2d ago
AI Research Intern - Speech and Audio Models
Posted 1mo agoBe an early applicant
What this role pays, in context
This listing does not state its pay. Data Scientist roles in India currently advertise a median of ₹24K a month.
25th pct ₹20.6K75th pct ₹27.1K
Median of 12 other live Data Scientist listings in India on this board that state pay. Roles that do not disclose a salary are excluded rather than counted as zero.
An experiment, not a product
A few thousand profiles now carry a sixty-second voice intro and we genuinely do not know what to do with them. This two-month internship is an open question rather than a roadmap item: what, if anything, can we responsibly learn from a candidate's audio clip.
The eight weeks
You transcribe a consented sample with an open speech recognition model, measure word error rate across Indian English accents because the published numbers will not hold here, then test whether transcript-derived features add anything to our existing profile matching. There is a fair chance the answer is no. A clean negative result, written up properly with the evidence behind it, counts as a successful internship on this team.
Terms and fit
The position is unpaid and runs at roughly twenty hours a week, two of those days in our Indiranagar office where you sit with the research group and get time set aside for reading. It is aimed at final-year students and recent graduates who have taken a signals or speech course, can work through a paper without being intimidated by it, and want to try research conditions before committing to a masters.
Responsibilities
- Transcribe a consented sample of candidate voice intros using an open speech recognition model
- Measure word error rate across Indian English accents and report where the off-the-shelf model breaks down
- Test whether transcript-derived features add anything to the existing profile matching signal
- Read and summarise three recent papers on speech representation learning for the research group
- Write up the result carefully, including a clean negative result if that is what the data supports
- Flag any privacy or consent concern you notice in how the audio is stored or used
Requirements
- Final-year student or recent graduate with coursework in signal processing, speech or machine learning
- Python and comfort with audio tooling such as librosa or torchaudio
- Able to read a research paper and extract the method from it
- Careful and honest about results, including the inconvenient ones
- Available for roughly twenty hours a week, two of them in the Indiranagar office
Skills
Benefits
- Certificate + Letter of Recommendation
- A seat with the research group and protected time for reading papers
- Honest exposure to research conditions before you commit to a masters programme
- Your write-up published on the internal research wiki with your name on it