About this role.
Synthesia is hiring a Senior Research Engineer to develop production-quality generative speech and voice-synthesis systems. The role focuses on streaming speech-to-speech models, expressive voice conditioning, post-training optimization, and low-latency inference. The engineer will work within a 40+ person R&D organization and contribute directly to voice models used in a commercial AI video platform. Success requires strong practical machine learning expertise, PyTorch proficiency, and the ability to rapidly prototype, evaluate, and productionize advanced audio-modeling approaches. Experience with speech generation, neural codecs, diffusion, LLMs, and distributed training is especially relevant.
Role DNA
A quick view of the complexity, pace, ownership and collaboration implied by the job description.
Job Complexity
5/5Pace & Pressure
4/5Autonomy Level
4/5Communication Load
4/5Salary analysis
Estimated compensation compared with the broader US market for similar roles.
Core skills
Skills and capabilities most closely associated with this opportunity.
Cover letter sample
Dear Synthesia Hiring Team,
I am excited to apply for the Senior Research Engineer - Voice role. My background in generative modeling, transformer-based systems, and end-to-end deep-learning development aligns closely with your work on expressive, real-time synthetic speech.
I would bring a practical research-engineering approach to prototyping, evaluating, and optimizing voice models, including distributed PyTorch training, post-training techniques, and latency-sensitive inference. I am particularly motivated by the opportunity to help deliver robust speech-to-speech and conversational voice capabilities at production scale.
I would welcome the opportunity to discuss how my experience can contribute to Synthesia's Audio Post-Training Team.
Sample interview questions
I would combine perceptual quality measures such as MOS or learned MOS predictors with streaming-specific measures: time to first audio, end-to-end turn latency, chunk-boundary artifacts, interruption handling, real-time factor, and stability under network or compute variation. I would also run human evaluation on conversational scenarios, because natural turn-taking and responsiveness often reveal failures not captured by offline metrics.
I would first profile the serving stack to identify whether latency is dominated by model compute, memory bandwidth, codec stages, or orchestration. Quantization is often the first option for reducing memory and accelerating supported kernels; distillation is appropriate when a smaller student can retain quality; pruning can help when the deployment runtime exploits sparsity effectively. I would validate each approach against quality, speaker similarity, prosody control, real-time factor, and production hardware constraints.
The controls must be well represented in training data, disentangled enough to avoid unwanted changes in linguistic content or identity, and consistently followed at inference. I would use structured conditioning inputs, targeted data balancing, control-specific evaluation sets, and preference or ranking objectives where appropriate. I would also test compositional controls, such as an emotional speaking style at a faster rate, to identify interaction failures.
I would use distributed data parallelism as a baseline and select sharding approaches such as FSDP when model or optimizer memory requires it. Important considerations include efficient data loading, mixed precision, gradient accumulation, checkpoint reliability, reproducibility, communication overhead, and monitoring training quality across workers. I would benchmark throughput and memory while ensuring that distributed behavior matches single-node validation results.
I would define explicit acceptance criteria covering speech naturalness, intelligibility, speaker similarity, controllability, streaming latency, cost, robustness, and maintainability. After reproducing a credible baseline, I would run controlled experiments against the current system on representative languages, voices, and conversational use cases. I would recommend integration only if the quality and operational gains justify the added serving complexity and can be sustained with reliable evaluation and monitoring.
Synthesia is the world’s leading AI video platform for business, used by over 90% of the Fortune 100. Founded in 2017, the company is headquartered in London, with offices and teams across Europe and the US.
As AI continues to shape the way we live and work, Synthesia develops products to enhance visual communication and enterprise skill development, helping people work better and stay at the center of successful organizations.
Following our recent Series E funding round, where we raised $200 million, our valuation stands at $4 billion. Our total funding exceeds $530 million from premier investors including Accel, NVentures (Nvidia’s VC arm), Kleiner Perkins, GV, and Evantic Capital, alongside the founders and operators of Stripe, Datadog, Miro, and Webflow.
What you’ll do at Synthesia
As a Research Engineer you will join a team of 40+ Researchers and Engineers within the R&D Department working on cutting-edge challenges in the Generative AI space, with a focus on creating high-quality, expressive and real-time synthetic voices. Within the team you’ll have the opportunity to work on the applied side of our research efforts and directly impact our solutions that are used worldwide by over 60,000 businesses.
If you are an expert in ML, LLMs, speech generation, conversational models, this is your chance to make a global impact. You will join our Audio Post-Training Team, which works on generative speech and voice synthesis, ensuring our in-house voice models reach production-level quality, speed, and robustness. Typical projects include:
Develop and evaluate streaming and speech-to-speech systems, enabling low-latency, interactive voice synthesis.
Adapt models for new conditioning inputs (emotion, speed, prosody, speaker control, etc.).
Implement post-training optimization techniques (quantization, pruning, distillation) to improve efficiency and latency in real-time speech generation.
Integrate and test novel architectures, such as neural codecs, diffusion, or flow-matching models, to enhance realism and responsiveness.
Contribute to defining new evaluation metrics for conversational speech, including latency-aware and online MOS prediction systems.
Stay updated with the latest research in audio diffusion, autoregressive models, neural codecs, and multimodal LLMs.
Apply DPO (Direct Preference Optimization) and distillation to fine-tune large-scale speech models.
What we’re looking for:
Strong understanding of generative modeling, ideally applied to sequential or multimodal data.
Hands-on experience with large language models (LLMs) or similar transformer-based architectures.
High proficiency in PyTorch, including experience with distributed training and model optimization.
Solid grasp of time-series modeling and tokenization, preferably in the context of audio or speech.
Demonstrated ability to prototype quickly, test hypotheses, and iterate efficiently.
Proven experience in training deep learning models end-to-end, from data preparation to evaluation.
Strong general software engineering skills, enabling contributions to a large, shared research infrastructure.
Nice to have experience:
Experience with real-time or streaming architectures is a big plus.
Familiarity with state-of-the-art architectures in audio and speech generation (e.g., diffusion models, neural codecs, flow-matching models, autoregressive decoders).
Experience with speech-to-speech or text-to-speech (TTS) systems.
Evidence of original research contributions, such as publications or open-source work in top-tier venues (e.g., ICASSP, Interspeech, NeurIPS, ICML).
Annual salary information is not provided for this position. Explore salary ranges for similar roles in our Salary Directory ›
Upload your resume now
- To unlock remote work opportunities and be discovered by global employers.
This job listing has been manually reviewed by the Jobicy Trust & Safety Team for compliance with our posting guidelines, including verification of the company's legitimacy, accuracy of job details, clarity of remote work policy, and absence of misleading or fraudulent content.
