Machine Learning Engineer - Voice AI & Generative Music (Part-time)

MWDN•Madrid, MD

About The Position

MWDN is seeking an Applied Machine Learning Engineer specializing in audio and generative AI for an AI-native music label. This company uses artificial intelligence to power its entire content-production process, including artist-consistent audio generation, vocal cloning, custom text-to-speech, and generative lyrics. They have developed an end-to-end music production pipeline. This role is a practical engineering position focused on turning sophisticated AI models into release-ready music products. The engineer will take ownership of the workflow, which includes audio data preparation, model training, vocal conditioning, inference, quality evaluation, and deployment.

Requirements

3+ years of experience training and deploying deep learning models in production, preferably within audio or NLP.
Hands-on experience in at least two of the following areas: Voice cloning or Singing Voice Conversion (SVC), Text-to-Speech model fine-tuning, LLM fine-tuning using techniques such as LoRA, DPO, and RAG.
Strong proficiency in PyTorch.
Practical experience with GPU cloud infrastructure, such as RunPod or AWS, as well as Docker and serverless inference.
A rigorous, evidence-based approach to model evaluation, including benchmarks, ablation studies, and blind testing.
Native or strong professional proficiency in Spanish, required for lyrics evaluation and voice quality assurance.

Nice To Haves

A music background or experience with music-production tools and workflows, including stems, MIDI, and DAWs.
Experience with singing-voice synthesis solutions such as ACE Studio, ACE-Step, RVC, so-vits-svc, or similar technologies.
Previous experience working with licensed celebrity or artist voices and consent-based voice AI.

Responsibilities

Fine-tune and improve an existing singing-voice conversion and cloning model, increasing its current 70–75% fidelity baseline to production-level quality of 90% or higher in blind listening tests.
Reproduce expressive vocal characteristics beyond basic timbre, including vibrato, falsetto, dynamics, and both spoken and sung delivery.
Evaluate different improvement strategies for singing voice models, including expanding the training dataset, testing alternative base models, and assessing enterprise APIs.
Take ownership of an existing validated LoRA fine-tune based on Chatterbox Multilingual for a Spanish-speaking voice.
Ensure accurate reproduction of a Mexican accent and correct pronunciation of Spanish sounds and letters.
Containerize the Text-to-Speech model and deploy it as a serverless inference endpoint using RunPod or a similar platform.
Integrate the Text-to-Speech endpoint with the existing web platform through an API.
Train a second version of the Text-to-Speech model using clean studio recordings to eliminate remaining output instability.
Develop a proprietary Spanish-language lyrics-generation model, with experience in regional Mexican genres considered a strong advantage.
Build the lyrics-generation solution using an open-weight LLM, LoRA fine-tuning, preference optimization through DPO, and RAG based on a curated content corpus.
Establish clear evaluation criteria and organize quality assessment for the lyrics model with native Spanish speakers.
Integrate the completed lyrics model into the existing frontend.