ML Engineer & Research Scientist, Speech (Full-time)

Location: San Francisco/Gurgaon · Type: Full-time

About Voice ArenaVoice Arena builds frontier evals, models and datasets for voice AI. Our mission is to make machines talk like humans, and to measure honestly how close they are.

We run public leaderboards for speech to text, text to speech, speech to speech and speaker diarization, with many more yet to come. Labs, voice platforms and enterprises use them to see how speech models perform on real speech, in real conditions and across languages, rather than on clean read audio.

Behind the boards is a research and engineering team that builds its own datasets, trains the models that keep its evaluations accurate and publishes its methods. What sets us apart:

- Research that gets published. Three papers at Interspeech 2026, where Voice Arena is also the Diamond Sponsor. Our methods go through peer review, not just a blog post.
- Partners across the field. We brought Hindi and Indian English to Hugging Face's Open ASR Leaderboard, and we partner with NVIDIA. Model providers send us new models for public and private evaluation.
- Data nobody else has. Monsoon, our dataset initiative for the Global South, is working towards 1,000 languages by 2027. Our accented English corpus covers 8 accents and 6,000 hours. This data gives us held-out test sets that models have not seen in training.
- A research community. Voice Research Club runs paper-first sessions for voice AI researchers, starting in San Francisco and New York, with more cities to follow.

About the roleOur mission is to make machines talk like humans. Before that is an engineering problem, it is a research problem: someone has to define what talking like a human actually means, measure how far today's models are from it, and show what closes the gap.

We're hiring an ML Engineer & Research Scientist to own that problem end to end. This role blends fundamental research and hands-on engineering, and we don't separate the two. You will decide which behaviours the best voice models should have, build the evals that measure them, define the data that teaches them, and train the models that prove it works on our leaderboards and others.

The core question for this role is easy: what does a speech model need to do to talk like a human, and how do we prove when it does?

This is a highly autonomous role. You'll work directly with the founders in a small, talent-dense team, with real ownership from day one. Everything you ship is public. When a new speech model launches, labs and enterprises check our boards to see where it stands, so your evals decide what gets measured and your models show the field what is possible.

In this role, you willThe work is one end-to-end research loop. You'll run it again and again, each time on a different open problem in speech:

1. Identify the problems in speech. Find the gaps between today's speech models and machines that talk like humans, across recognition, synthesis, turn-taking, conversation, emotion, languages and accents. Listen to where models fail, read the literature, talk to the labs and users who hit these failures every day,



and turn what you find into clear problem statements.
2. Define the behaviour we want models to have. For each problem, write down precisely how a model should behave: what a good response sounds like, what counts as a failure, and where the line between them sits. This becomes the target labs build towards.
3. Design the evals. Turn that behaviour into a measurement: test sets, automatic metrics, and human evaluation protocols where metrics fall short. Run current models through it and show where they stand and where they break.
4. Define the datasets. Specify the data that would teach models the missing behaviour: what to record, from whom, in which languages and conditions, and how to label it. Our operations team builds it to your spec, and you own its quality.
5. Train models on that data. Train and fine-tune models on the new data to prove it works. You'll be responsible both for making the numbers go up and for making sure the numbers mean something. The bar: models that climb our evals, our leaderboards and other public leaderboards.
6. Publish and open-source. Write up the problem, the eval, the data and the results, and release the models. We want this work to be cited and built on, not just shipped.

In a typical week, you might:
- Listen to hundreds of model outputs to find a failure that no metric is catching yet
- Write the spec for a new eval and review the first batch of test data with the operations team
- Run training runs and ablations, and dig into results that don't quite make sense
- Check a new model launch against our boards and explain what its scores really mean
- Talk to researchers at labs about the problems they are stuck on
- Write up results for a paper, a blog post or a leaderboard release

What matters mostEverything in the loop serves two outcomes. They are how this role will be measured.

1. Building the right evals An eval only matters if climbing it means real progress. Many speech benchmarks saturate, or reward models that learn the test instead of the skill. You'll own making sure a better score on our evals means a model is genuinely closer to the frontier of voice AI and to machines that talk like humans. You should be able to say what an eval measures, what it misses, and how it could be gamed. In practice:

- Start from behaviour. Every eval begins with a behaviour we want models to have, not with a metric that happens to be easy to compute.
- Check against people. Validate that scores agree with what human listeners actually prefer, and use human evaluation where automatic metrics fall short.
- Keep it hard to game. Use held-out and refreshed test sets, watch for contamination, and retire an eval once it stops separating good models from great ones.
- Cover real speech. Test across languages, accents, code-mixing, noise and real conversation,



not only clean, read English.
- Be rigorous. Report confidence intervals, and tell real effects apart from noise and bugs before calling a ranking.

2. Building the models Evals tell the field where to go; models prove the path works. You'll train models on the data you defined and show that they climb, so the behaviours we set and the data we design are backed by results, not opinion. In practice:

- Train and fine-tune. Work with open ASR, TTS and speech to speech models, fine-tuning them on the datasets built to your spec.
- Prove what caused the gain. Run clean ablations against honest baselines, so the improvement is shown to come from the new data and behaviour, not from tuning to the test.
- Climb public boards. Move up our leaderboards and other existing ones, such as Hugging Face's Open ASR Leaderboard, with results others can reproduce.
- Train efficiently. Keep training runs reproducible and make the most of every GPU hour.
- Release openly. Open-source the models with clear model cards, so labs and researchers can build on them.

What we offer

- Hard, open problems. Own the research questions that decide whether machines can talk like humans, across recognition, synthesis and conversation.
- Open-source models. Release the models you train, under your name.
- Co-authorship on papers. Co-author our papers at Interspeech, ICASSP and other top venues.
- Real impact. Your evals shape how labs build speech models and how the industry chooses between them.

Minimum qualifications

- An MS or PhD in speech, ML or NLP, or an equivalent industry research record
- Strong Python and PyTorch, and comfort with large audio datasets
- Statistical care: you know when a ranking difference is real and when it is noise
- Informed opinions about where voice AI is heading, not just familiarity with the tools
- Solid engineering habits: version control, tests, reproducible experiments and pipelines others can run

Preferred qualificationsYou don't need to match every bullet. Apply even if you meet only some of these.

- Papers at Interspeech, ICASSP, ACL, NeurIPS or similar
- Experience with multilingual, low-resource or code-mixed speech
- Work at a speech lab or voice AI company
- A benchmark, dataset or open-source tool others in the field use
- Experience running GPU training or inference at scale
- Experience designing human evaluation or annotation studies

Why join Voice Arena

- Shape how voice AI gets built. Labs, voice platforms and enterprises use our boards to decide which models to build, ship and buy.
- See every model first. Work with the newest speech models across languages, often before most of the field has tried them.
- Publish and be seen. Your work goes to Interspeech and ICASSP, to Hugging Face, and out to the voice AI community on X and LinkedIn.
- Measure the whole world's speech. Our data and evals reach far beyond English, across the wider Global South.
- Join early. Join a small, talent-dense team early, and shape the product, the team and the research agenda.

How to applySend your CV, a short note and work you're proud of to [email protected].

📌 ML Engineer & Research Scientist, Speech (Full-time) (Gurugram)
🏢 Voice Arena
📍 Gurugram