Job Description
About The Role:
As a Deep Learning Engineer (Voice), you will own the end-to-end life cycle of our voice model training — from large-scale data curation and foundational training to high-precision fine-tuning. We are looking for an engineer who can dive deep into model architectures to debug and optimize performance at the structural level. Your primary mission is to achieve world-class accuracy by ensuring our models encompass the full spectrum of vocal, linguistic, and acoustic diversity, including performance across varied real-world environments.
What you’ll be Responsible for:
Tackle a wide spectrum of audio challenges—including TTS, ASR, Diarization, Denoising, and Turn Detection—applying specialized deep learning techniques to each as project needs evolve.
Design and optimize models across the complexity spectrum, from lightweight, latency-sensitive specialized models to large-scale foundational architectures.
Adopt open-source codebases, leverage specialized toolkits, or implement state-of-the-art research papers from scratch based on specific requirements.
Own the entire model lifecycle, including massive-scale data creation/curation, foundational training, and high-precision fine-tuning to meet production standards.
Maintain detailed daily documentation of experiments and progress. Take an active, vocal role in everyday technical discussions to ensure engineering efforts are perfectly aligned with team expectations and project milestones.
What we are looking for, in you:
Must have:
Deep theoretical and practical understanding of Machine Learning and Deep Learning architectures (Transformers, CNNs, RNNs, etc.) and their application to complex audio problems.
Strong grasp of fundamental speech processing concepts, including acoustic modeling, feature extraction (MFCCs, Mel-spectrograms), and neural vocoders.
Ability to deconstruct model architectures to identify bottlenecks, debug training instabilities, and optimize for both accuracy and performance.
Optimizing the voice model architecture and customizing it to achieve the desired outputs.
Good to have:
Familiarity with digital signal processing (DSP) specifically for speech, such as noise reduction, echo cancellation, and spatial audio techniques.
Hands-on experience with advanced training optimization libraries and frameworks like Ray or Unsloth.
Basic performance profiling using frameworks like triton and vLLM
Knowledge of leading vendor models (such as ElevenLabs) and the ability to benchmark custom-built models against them.
Required:
5 - 7 years of industrial experience working in voice AI
BE/BTech/ME/MTech/PhD with background in artificial intelligence
(degree branch does not matter)
Proficiency in Python and the modern ML stack (PyTorch, NumPy, Librosa).
Should have worked on speech model architectures like: Whisper, Parakeet, VITS, Qwen TTS, Noise Cancellation, Noise removal, etc (or anything relevant),
Experience with Linux environments, shell scripting, and version control (Git).
Knowledge of model serving frameworks and related modules such as vLLM, Triton, TensorRT, and TensorRT-LLM.
Implement end-to-end pipelines including text normalization, G2P mapping, NLP intent extraction, and emotion/prosody control.
Knowledge of standards in ML modeling pipeline: audio cleaning, experiment tracking, quality assessment, automatic checkpointing etc.
Basic performance analysis of model
Why join us?
Impactful Work:
Play a pivotal role in safeguarding Tanla's assets, data, and reputation in the industry.
Tremendous Growth Opportunities : Be part of a rapidly growing company in the telecom and CPaaS space, with opportunities for professional development.
Innovative Environment:
Work alongside a world-class team in a challenging and fun environment, where innovation is celebrated.
Tanla is an equal opportunity employer. We champion diversity and are committed to creating an inclusive environment for all employees.
www.tanla.com
