Speech AI Engineer

Location & work mode

This role is based in Bengaluru, India, and works from our office. We care about focused time and short feedback loops, which is easier in a room together while the team is this size.

What you'll do

• Train and fine-tune text-to-speech and speech-to-speech models, from data through to a checkpoint that ships.
• Own the quality bar for a language or a voice family: what good sounds like, how it is measured, and what has to change to get there.
• Work on the parts that make speech feel live — streaming inference, time to first audio, and the tradeoffs between latency and quality.
• Build the evaluation that tells us we improved: objective metrics, listening tests, and the tooling around both.
• Take on the hard cases directly — code-switching, low-resource languages, expressive delivery, and voice cloning.

What we're looking for

• Experience training speech or audio models end to end, in PyTorch, and getting them to a state someone else can run.
• A working understanding of modern TTS: neural codecs, autoregressive and diffusion approaches, vocoders, and where each breaks.
• Comfort with the unglamorous half — data collection, cleaning, alignment, and normalization, which is usually where the quality actually comes from.
• The judgement to tell a real improvement from a metric that moved.
• Clear writing. We work from written arguments more than meetings.

Bonus

• You have worked on Indic or African languages, or any low-resource setting.
• You have shipped speech-to-speech or real-time voice systems.
• You have published, or contributed to open-source speech work.
• You speak more than one of the languages we support, and can hear when a voice is wrong.

What we offer

• Competitive salary, discussed early in the process.
• Direct ownership of a system in production, not a slice of a backlog.
• A small team and a short path from idea to shipped.
• Hardware and compute you need to do the work.
• Based in our Bengaluru office.