Research Engineer - Speech Models

Vaani Research Labs; Bengaluru; On-site; INR 1,500,000 - 2,100,000; Full time

Job summary

Vaani Research Labs is building the Latent Speech Model, a foundation model powering all offerings. You will join a six-person research team and see results go live on real customer calls within weeks, not quarters.

You will work on evidential deep learning pre-training for speech, develop an acoustic tokenizer, and tackle code-mix robustness across 22 languages with low-resource phonetics. Collaboration with customers is expected.

Qualifications

• Strong grounding in speech, audio, or sequence modelling.
• Fluent in PyTorch and able to translate maths from papers into training code.
• Willingness to talk to customers and explain results.

Responsibilities

• Train and evaluate speech models using evidential deep learning pre-training.
• Work on the continuous VAE-based acoustic tokenizer and ablations for discrete vs continuous representations.
• Improve code-mix robustness across 22 scheduled languages with focus on mid-sentence switches and low-resource phonetics.
• Own benchmark methodology (WER, MCD, TTFT) and defend results publicly.
• Contribute to the Latent Speech Model paper and Lab Notes as results are shareable.

Skills

Speech modelling

PyTorch

Mathematics reading

Customer communication

Job description

You will work on the Latent Speech Model - the foundation model underneath everything Vaani ships. This is not fine-tuning someone elses checkpoint. We train our own architecture on a 4.3 lakh hour proprietary Indic corpus, and the research questions are genuinely open.

You will sit inside a six-person research team, and your work reaches production within weeks, not quarters. If a result holds, it goes live on real customer calls.

What Youll Do

Modelling

• Train and evaluate speech models using our evidential deep learning pre-training approach - uncertainty quantification is a first-class objective, not an afterthought
• Work on the continuous VAE-based acoustic tokenizer and push on the discrete-versus-continuous question with actual ablations
• Improve code-mix robustness across 22 scheduled languages, where the hard problems are mid-sentence switches and low-resource phonetics

Evaluation & publication

• Own benchmark methodology - WER, MCD, TTFT - and defend the numbers publicly
• Contribute to the LSM paper targeted for publication, and to Lab Notes as results become shareable

What Were Looking For

• Strong grounding in speech, audio, or sequence modelling - a thesis, a paper, or shipped work all count equally
• Fluent in PyTorch and comfortable reading maths in a paper and turning it into working training code
• Willingness to talk to customers. Researchers here hear the calls their models handle
• Bonus: experience with low-resource languages, TTS/ASR training pipelines, or large-scale audio data engineering