Speech/audio ML researcher at Mila: neural speech codecs and LLM continued pretraining, with very high recent output
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
26
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—
Output is not scored yet: the per-work volume and venue data needed for this axis are insufficient in the current snapshot. It would measure publication volume and consistency over time; an h-index of 26 exists, but no output score has been computed.
Impact—
Impact is not scored yet: beyond the reported h-index (26) and i10-index (45), the citation distribution and uptake data needed for this axis are unavailable. It would measure how widely and how durably the works are cited and adopted.
Focus0/100
The focus score is 0, the lowest on this radar: the field distribution splits evenly (0.5 natural language processing, 0.5 speech processing / neural audio coding), and topics span five areas — speech recognition and synthesis (121 works), speech and audio processing (100), music and audio processing (59), speech and dialogue systems (35), and topic modeling (20). Breadth, not concentration, defines this record.
Collaboration36/100
Collaboration scores 36, moderate-low: the snapshot records only 5 distinct coauthors, with repeated work concentrated on Cem Subakan, Fırat Öncel, Salman Hussain Ali, Çağatay Yıldız, and Luca Della Libera. On the two reviewed 2026 papers Ravanelli is neither first nor corresponding author (middle on the LLM pretraining study, last on ZipCodec), consistent with a senior role inside a compact, stable coauthor group.
Activity100/100
Activity is the strongest axis at 100: 71 works in the most recent three years, and the yearly series climbs from a single work in 2012 to 13 in 2021, 17 in 2023, a peak of 28 in 2024, 19 in 2025, and 24 already recorded in 2026 — a pace roughly triple the mid-2010s baseline of 4–9 works per year.
Reproducibility55/100
Reproducibility scores 55, computed from 2 reviewed papers with a mean readiness of 55/100; both carry a medium rating. Evolve finds the continued-pretraining study credible but notes its gains are net rather than absolute, while ZipCodec bundles several changes without the ablations needed to isolate each contribution.
Research summary
Mirco Ravanelli is a researcher at Mila – Quebec Artificial Intelligence Institute whose OpenAlex-derived snapshot splits his work evenly between natural language processing (0.5) and speech processing / neural audio coding (0.5). The bibliometric record is long and productive: it opens with a single work in 2012 and climbs to 28 works in 2024, 19 in 2025, and 24 already recorded in 2026, with 71 works across the most recent three years — the basis for a perfect activity score of 100. Citation aggregates show an h-index of 26 and an i10-index of 45. Topic analysis places speech recognition and synthesis (121 works) and speech and audio processing (100) at the core, with music and audio processing (59), speech and dialogue systems (35), and topic modeling (20) alongside. Two 2026 arXiv papers reviewed by Evolve mark the current agenda: time-incremental continued pretraining adds post-cutoff knowledge to LLMs at near-zero forgetting cost, though gains are net rather than absolute, while ZipCodec claims the first streaming codec at 6.25 Hz and 0.80 kbps, with broad, credible results but no ablations for its bundled changes. Both are rated medium for reproducibility readiness (mean 55 over two papers); the focus score of 0 reflects the evenly split field distribution, and collaboration (36) rests on a compact core of frequent coauthors including Cem Subakan, Fırat Öncel, Salman Hussain Ali, Çağatay Yıldız, and Luca Della Libera.
Collaboration pattern
Collaboration is the moderate-low axis (36): only 5 distinct coauthors appear in the snapshot, and repeated work clusters on five names — Cem Subakan, Fırat Öncel, Salman Hussain Ali, Çağatay Yıldız, and Luca Della Libera. On the two reviewed 2026 papers Ravanelli is neither first nor corresponding author (middle on the LLM pretraining study, last on ZipCodec), consistent with a senior-collaborator role within a compact, stable coauthor group rather than a broad network.
Research keywords
Highlights
ZipCodec: streaming neural codec claimed at 6.25 Hz / 0.80 kbps
Time-incremental continued pretraining updates LLM knowledge without catastrophic forgetting
Timeline
- 2012First work recorded in the OpenAlex snapshot (1 publication)
- 2021Annual output reaches 13 works, an early high before the 2023–2026 acceleration
- 2024Peak annual output in the snapshot: 28 works recorded
- 202519 works recorded, keeping the pace near the 2024 peak
- 202624 works already recorded; both Evolve-reviewed papers appear — time-incremental continued pretraining of LLMs and ZipCodec streaming speech coding
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.
- Bibliometrics (h-index, i10-index, yearly counts, topics, field distribution) come from an OpenAlex-derived snapshot; identity is resolved by OpenAlex author disambiguation only.
- The output and impact dimensions are not scored in this radar; only raw citation aggregates and yearly counts are available.
- The reproducibility score rests on only 2 reviewed papers, so it is indicative rather than representative of the full record.
- The snapshot lists the country as 'US' alongside Mila – Quebec Artificial Intelligence Institute; treat the country field as uncertain metadata.