ML Research Engineer, Speech

Company and team

Location and team

• Location: Bengaluru; the posting describes the work arrangement as on-site / hybrid.
• The role is on the Speech Research team and reports into the CTO organization.

Role overview

Responsibilities

• Pre-train STT models from scratch and fine-tune them for Indian languages, Hinglish, and BFSI vocabulary.
• Train TTS models from scratch and extend them to new voices, languages, and speaking styles.
• Build speech-to-speech components, including audio tokenizers and codecs, speech encoders connected to LLMs, and streaming decoders.
• Implement research ideas in PyTorch, including model architectures, tokenizers, and training loops initialized from scratch.
• Build data pipelines for cleaning, segmentation, forced alignment, pseudo-labeling, and augmentation, including noise, codec, and 8 kHz telephony simulations.
• Maintain evaluation suites for WER/CER, entity error rate, code-switch accuracy, MOS, speaker similarity, and latency.
• Optimize serving with quantization, streaming chunking, ONNX/TensorRT export, and GPU batching.
• Analyze production errors, create test sets, and work with the platform team to close the loop.
• Write experiment reports and share findings in weekly research reviews.

Experience and education

• 1–3 years in ML, with at least one real speech or audio project from industry, a research lab, or a strong thesis.
• B.Tech, M.Tech, or MS in computer science, electrical engineering, or a related field; a PhD is not required.
• Strong Python and PyTorch; able to implement model architectures and training loops from scratch.
• Working knowledge of spectrograms, MFCC/mel features, CTC, attention-based seq2seq, and vocoders.
• Hands-on with at least one of NeMo, ESPnet, Hugging Face Transformers/Audio, fairseq, Coqui, or k2/icefall.
• Trained at least one model from random initialization, including multi-GPU jobs, mixed precision, and W&B or MLflow experiment tracking.
• Careful evaluation practices, including awareness of why WER can mislead and the importance of checking data.

Preferred experience

• Speaking or understanding Hindi or another Indian language is a plus.
• Speech publications, benchmark results, or open-source contributions are nice to have.
• Exposure to WebRTC, SIP, streaming inference, or telephony audio is nice to have.
• Familiarity with LLM fine-tuning (LoRA, SFT) or audio-language models is nice to have.

Success measures