ML Research Engineer - Speech

Carrerlift; Bengaluru; In loco; INR 1.200.000 - 2.200.000; Tempo pieno

Descrizione del lavoro

Carrerlift in Bengaluru is hiring a research engineer to develop speech and audio ML systems from ground up. You will pre-train and fine-tune models, focusing on Indian languages and Hinglish, with a strong emphasis on PyTorch implementations.

Ideal candidates have 1–3 years ML experience, hands-on PyTorch training from random initialization, and multi-GPU / mixed-precision expertise. This role is based in Bangalore with opportunities to shape production-ready pipelines.

Competenze

• At least one practical speech or audio ML project.
• Hands-on experience training a model from random initialization in PyTorch.
• Experience with multi-GPU model training.
• Experience with mixed-precision training.

Mansioni

• Pre-train Speech-to-Text models from random initialization.
• Train Text-to-Speech models from scratch using modern speech architectures.
• Work with architectures like Conformer, Zipformer, Flow-Matching.
• Fine-tune speech systems for Indian languages and Hinglish.
• Develop audio tokenizers and components connecting speech systems with LLMs.
• Build low-latency streaming decoders for real-time speech applications.
• Translate research ideas into clean PyTorch implementations.
• Build and manage speech-data pipelines (8 kHz telephony sim, forced alignment).
• Optimize models for production using ONNX, TensorRT, and quantization.

Conoscenze

Speech/Audio ML

PyTorch

Multi-GPU training

Mixed-precision training

Strumenti

PyTorch

Descrizione del lavoro

About the role:

• The opportunity focuses on research engineering for speech and audio machine-learning systems.
• Candidates should have 1–3 years of machine-learning experience.
• The role involves developing speech models from the ground up rather than only adapting existing systems.
• Applicants are expected to be currently based in Bangalore.

What you'll work on:

• Pre-train Speech-to-Text models from random initialization.
• Train Text-to-Speech models from scratch using modern speech architectures.
• Work with architectures and techniques such as Conformer, Zipformer, Flow-Matching and Neural Codecs.
• Fine-tune and adapt speech systems for Indian languages and Hinglish.
• Optimize models for specialized BFSI terminology and vocabulary.
• Develop audio tokenizers and components that connect speech systems with large language models.
• Build low-latency streaming decoders for real-time speech applications.
• Translate research-paper ideas into clean and functional PyTorch implementations.
• Build and manage speech-data pipelines, including 8 kHz telephony simulation and forced alignment.
• Optimize models for production using ONNX, TensorRT and quantization techniques.

What we're looking for:

• Applicants must have at least one practical speech or audio machine-learning project.
• Hands-on experience training a model from random initialization in PyTorch is required.
• Candidates should have worked with multi-GPU model training.
• Experience using mixed-precision training is also required.

Note:

Candidates should have 1–3 years of ML experience, at least one real speech/audio project, and must currently be in Bangalore. The post also allows interested candidates to reach out by DM.