Mila researcher at the intersection of neural speech coding and continual pretraining for language models
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
12
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
This axis would measure publication volume and productivity over time. The snapshot supplies a recent-3-years count of 57 works, but no complete, verified output record, so Evolve does not score overall output for this researcher yet.
Impact—
This axis would measure citation-based influence such as citations per paper and field-weighted impact. Although the snapshot lists h-index 12 and i10-index 13, the fuller citation record needed to score impact is not available.
Focus0/100
Focus scores 0 because the portfolio is deliberately split: the field distribution is exactly 50% natural language processing and 50% speech processing / neural audio coding. Topic counts confirm breadth across at least five clusters — Speech Recognition and Synthesis (49), Music and Audio Processing (46), Speech and Audio Processing (37), Domain Adaptation and Few-Shot Learning (13), and Topic Modeling (9) — so no single dominant line emerges.
Collaboration36/100
Collaboration scores 36 on the strength of 5 distinct coauthors recorded in the snapshot. The recurring partners are Mirco Ravanelli, Fırat Öncel, Salman Hussain Ali, Çağatay Yıldız, and Luca Della Libera. Notably, both 2026 catalog entries list him in a middle author position and not as corresponding author, suggesting contribution within larger team-led papers rather than small self-led dyads.
Activity100/100
Activity scores 100, the maximum. The counts-by-year series shows sustained output that accelerates sharply late: 1 work in 2014, 3 in 2017, 9 in 2023, 20 in 2024, 17 in 2025, and 20 in 2026, with 57 works across the most recent three years. The 2024–2026 run is the densest stretch in the record and anchors the top score.
Reproducibility55/100
Reproducibility scores 55, the mean readiness across the 2 papers Evolve has assessed. Both entries — the time-incremental continued-pretraining study and ZipCodec — are rated medium reproducibility: credible multi-family and broad results respectively, but with stated gaps such as missing ablations for ZipCodec's bundled changes and net rather than absolute gains for the pretraining study.
Research summary
Cem Subakan is a researcher at Mila — Quebec Artificial Intelligence Institute in Canada, profiled from OpenAlex bibliographic data. His recorded work splits evenly across two fields: natural language processing and speech processing / neural audio coding, each at 50%. Topic frequencies reinforce the audio-centric core — Speech Recognition and Synthesis (49), Music and Audio Processing (46), and Speech and Audio Processing (37) — with smaller clusters in Domain Adaptation and Few-Shot Learning (13) and Topic Modeling (9). Activity is the standout signal: counts by year rise from 9 works in 2023 to 20 in 2024, 17 in 2025, and 20 in 2026, totaling 57 works over the most recent three years. The snapshot reports an h-index of 12 and an i10-index of 13. In 2026 Evolve reviewed two of his arXiv papers, both with him in a middle author position: ZipCodec, claimed as the first streaming codec at 6.25 Hz and 0.80 kbps (results judged broad and credible, though bundled changes lack ablations), and a time-incremental continued-pretraining study that adds post-cutoff knowledge to LLMs at near-zero forgetting cost (gains are net, not absolute). Reproducibility readiness is medium for both.
Collaboration pattern
Cem Subakan's recorded collaboration footprint is compact — 5 distinct coauthors in this snapshot, yielding a collaboration score of 36. His most frequent partners are Mirco Ravanelli, Fırat Öncel, Salman Hussain Ali, Çağatay Yıldız, and Luca Della Libera. Both 2026 catalog papers place him as a middle author rather than corresponding author, consistent with contributing to team-scale systems work rather than leading small two-author studies.
Research keywords
Highlights
ZipCodec: streaming speech coding at 6.25 Hz and 0.80 kbps
LLM knowledge updates without catastrophic forgetting
A two-track 2026 output: audio codecs and LLM continual learning
The 50/50 field split and the 2026 activity peak (20 works) culminate in two arXiv entries reviewed by Evolve — one in speech processing / neural audio coding and one in natural language processing — mirroring the researcher's evenly divided topic profile.
Timeline
- 2014Earliest year in the snapshot's counts-by-year record, with 1 work.
- 2017An early cluster of 3 works appears in the counts-by-year series.
- 2023Output steps up to 9 works, the start of a steep climb.
- 2024Annual output reaches 20 works, the level later matched in 2026.
- 202620 works, including two arXiv entries reviewed by Evolve: ZipCodec and the time-incremental continued-pretraining study.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric figures (h-index, i10-index, counts-by-year, topics, coauthors) come from an OpenAlex-based snapshot only.
- Output and impact axes are unscored because the snapshot lacks complete publication and citation records.
- Evolve has reviewed only 2 papers for this researcher; verdicts apply to those entries alone.
- Counts-by-year has gaps (e.g., no recorded works for 2015–2016 and 2018), which may reflect database coverage rather than true inactivity.