The problemOff-the-shelf speech models are trained for clean English. Our users speak dozens of African and Middle Eastern dialects, over noisy lines, code-switching mid-sentence. The gap between a model that works in the demo and one that works for them is the whole job. You'll own the fine-tuning that closes it: taking base speech and language models and making them actually understand, and sound natural to, the people we serve.

The problemOff-the-shelf speech models are trained for clean English. Our users speak dozens of African and Middle Eastern dialects, over noisy lines, code-switching mid-sentence. The gap between a model that works in the demo and one that works for them is the whole job. You'll own the fine-tuning that closes it: taking base speech and language models and making them actually understand, and sound natural to, the people we serve.

Why it's hard hereThis is low-resource, messy-data territory. For a lot of the dialects we care about, the clean training corpus you'd want doesn't exist, so you'll be building it: sourcing, cleaning, labeling, and augmenting audio, then running the fine-tunes and proving they moved the numbers that matter. Word error rate on a real accent, not a benchmark. Naturalness a native speaker would actually sign off on. You'll live in the loop between data, training runs, and evaluation.

What You'll Own

Training/ fine-tuning STT and TTS (and where relevant,



the language layer) for our target dialects and acoustic conditions

The training data pipeline: sourcing, cleaning, labeling, and augmentation for low-resource languages

Evaluation you can trust: WER on real-world audio, naturalness, robustness to noise and code-switching

Feeding real production performance back into the next round of training

What We're Looking For

Strong applied ML experience, with models you've fine-tuned and shipped to production

Hands-on with speech: ASR/STT and/or TTS, and the training stack around them (PyTorch, etc.)

Real experience with the unglamorous parts: data collection, cleaning, and labeling for messy audio

Rigor around evaluation, especially where standard benchmarks don't reflect your users

Comfort owning a model end to end and shipping fast with imperfect data

Nice to have

Low-resource, multilingual, or multi-dialect speech work

Experience with modern speech architectures (Whisper-family, speech-to-speech, neural TTS)

A feel for the latency and size constraints of serving these models in real time

Genuine interest in emerging markets or speech tech

The setupYou’ll work directly with our CTO on a team of ~10 that ships constantly. High ownership, short path to decisions, real users the day you ship. London-based, in-office.

#J-18808-Ljbffr

📌 Voice AI Engineer (Madrid)
🏢 AethexAI
📍 Madrid