Full-sphere emission and reception directivities of seven echolocating bat species
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "Full-sphere emission and reception directivities of seven echolocating bat species".
Dev: Here's the gist of the paper. Echolocating bats shape their calls and the returning echoes with their noseleaf or mouth, head and pinnae Understanding these spectrospatial directivity patterns is central to bat echolocation.
Rosa: Dataset Description The paper describes a dataset of simulated far-field emission and reception directivities of seven echolocating bat species from three families (Phyllostomidae, Rhinolophidae and Vespertilionidae).
Dev: For each species, the dataset contains several components including (1) a closed, simulation-ready head mesh derived from a micro-computed tomography (micro-CT) scan, (3) the complexvalued far-field transfer functions of both ears and of the emitter on the full sphere from 20 to 160 kHz, and (5) a skeleton for the pinnae and noseleaf, with which the meshes can be posed and simulated again.
Rosa: The emission patterns of the five nasal emitters are BEM results, whereas the emitters of the two mouth-emitting vespertilionids are analytical piston models calibrated to published beam widths.
Paper summary: Rosa: So we’re looking at this paper called "Full-sphere emission and reception directivities of seven echolocating bat species." Basically, the authors built a dataset of simulated far-field emission and reception directivities for seven different bat species from three families: Phyllostomidae, Rhinolophidae, and Vespertilionidae. They’re showing how the noseleaf or mouth shapes these calls and echoes.
Dev: That sounds like a big undertaking because they are providing head meshes made from micro-computed tomography scans for each species. They are giving us the complex valued far-field transfer functions of both ears and the emitter on the full sphere, covering frequencies from twenty to one sixty kilohertz in steps of zero point five kilohertz.
Taro: So, what’s the main point here? Is it just cataloging bat anatomy or is there something deeper about how this shapes their echolocation patterns?
Rosa: It’s about understanding those spectrospatial directivity patterns because that’s central to how bats actually use echolocation. The paper claims this dataset is useful for simulating bat biosonar and echo acoustics, as well as for localization and head movement models.
Dev: And they provide a set of five nasal emitters using boundary element method results, but the two mouth emitting vespertilionids have analytical piston models calibrated to published beam widths. They even include skeletons for the pinnae and noseleaf so you can pose them and simulate again how posture shapes those directivities.
Taro: If we can pose them, that means we can test what happens when a bat changes its position, which is crucial for real-world scenarios where the world misbehaves. But what about the limitations of this data? Does it capture everything?
Rosa: The paper notes that every species is represented by only one preserved specimen, so they are actually a collection of specimen based reference models, not averages of species. They also state that these responses are for acoustically rigid heads and not full solutions of the near field.
Dev: That’s important because it means we can use this data to study morphology and directivity comparisons, or as training data for acoustic learning, but we shouldn't treat these as perfect measurements of what a bat actually experiences in flight.
Taro: I wonder if the way they derived the source placements helps us understand where bats are actually emitting sound in complex environments.
Rosa: They used a rule based on maximizing the standardized sum of front back ratio and occlusion for the other five species, which is how they placed those nasal emitters. And there’s a sensitivity analysis showing that alternative ear positions correlate with the nominal one with medians of zero point seven five to one point zero zero.
Dev: That suggests that even if you change where you put your ears on this model, the overall pattern correlation stays fairly consistent, which is a good sign for robustness in our simulations. But they also mention how sensitive the directivity index is to the radius of the piston emitters, changing by a median of one point five to one point nine decibels when that radius changes by twenty percent.
Paper summary: Taro: So if we are designing biomimetic sonar, we have to be careful about how much we rely on those specific source assumptions and parameters. What about comparing these simulated results to actual measurements?
Rosa: They did technical validation by comparing the solver against an analytical solution and independent data, showing consistency checks with reference ERTFs of P. discolor and M. microtis, for example. They also reproduce a measured HRTF just like Mesh2HRTF does, meaning both their simulations deviate from real measurements in the same way.
Dev: That’s interesting because they found that to align the measurement directions with the mesh, you need an additional rotation of about twenty six degrees to get those median pattern correlations up from zero point seven zero to zero point eight two.
Taro: So what does this tell us practically for someone listening who isn't a bat expert? It seems like it’s confirming that the physics of the head shape really dictates how sound travels, even when you change the input parameters slightly.
Rosa: Exactly, because what we’re seeing is that the geometry—the noseleaf or mouth—is directly shaping those spectrospatial patterns. The whole point of this paper is to give us a concrete way to model and predict bat echolocation behavior based on their physical structure.
Dev: And they’ve given us the tools, including the code that computes the responses and a small simulator that turns them into binaural echo signals, which is pretty useful for putting these directivity calculations into practice.
Taro: It feels like a solid foundation for building more complex models of bat biosonar, but I still see limitations because they explicitly state that the data represents rigid heads and fixed ear postures.
Rosa: That’s true; the limitation they flag is that since it only represents rigid heads and fixed ear postures, it doesn't capture variation within a species or how a bat might move its head dynamically during echolocation.
Dev: And another practical point from the paper is that for mouth emitters like Eptesicus fuscus and Myotis dau, the nominal oral emitters are calibrated to search calls in free flight in the open field, which means users need to adapt their beam width based on what situation they are modeling.
Taro: So, while this dataset gives us a strong starting point for simulating directivity based on morphology, we have to remember that the fidelity depends entirely on how well we model those rigid constraints and fixed postures.
Rosa: Right. So this paper is essentially giving researchers a comprehensive set of simulated far-field data for seven species, which they’ve modeled with high resolution up to one hundred sixty kilohertz. This sets a benchmark for simulating bat acoustics based on physical anatomy.
Conclusion: Rosa: So we've been looking at how researchers built this dataset of simulated far-field emission and reception directivities for seven bat species, which is what this paper is about, "Full-sphere emission and reception directivities of seven echolocating bat species."
Dev: Right. Basically, they took those head scans and the ear transfer functions from those different families—Phyllostomidae, Rhinolophidae, Vespertilionidae—and put them all together onto a full sphere model.
Taro: So the main idea is that they are trying to map out exactly how much sound a bat throws out in every direction and how well it hears echoes coming back from those directions.
Rosa: Exactly. It’s about seeing the entire pattern, not just one part of it, which is super important for understanding how bats actually use sonar in the real world.
Dev: And they’re using these models to figure out if a bat's head shape and its ears are optimized for searching in free space or something else.
Taro: It gives us a way to test those assumptions about where the sound is coming from, you know, when we try to model things that aren't just sitting still in a perfect vacuum.
Rosa: It helps bridge the gap between what we see on a head scan and what’s actually happening when that bat is flying around.
Dev: So, this work sets up a kind of benchmark for simulating bat acoustics based on their actual physical anatomy, which is pretty cool for testing new models.
Taro: It’s interesting because they're showing how sensitive the results are to the assumptions they make about where the sound source is exactly located on that head.
Rosa: Yeah, and while it’s a powerful tool for building those simulations, we have to remember it only works for these rigid heads and fixed ear postures.
Dev: That's the catch; it doesn't account for how a bat actually moves its head or changes its posture during echolocation calls in real flight.
Taro: Exactly, so the next step is figuring out how to move this model beyond being just a static representation of anatomy.
Jan Steckel, Flanders Make
Cosys-Lab · University of Antwerp
eess.AS
Submitted: 2026-11-09
Updated: 2026-11-09
Comments: 34 pages, 8 figures, supplementary information appended. Data: https://doi.org/10.5281/zenodo.23729269. Code: https://github.com/Cosys-Lab/bat_ertf_dataset
Code: https://github.com/Cosys-Lab/bat_ertf_dataset
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Echolocating bats shape their calls and the returning echoes with their noseleaf or mouth, head and pinnae Understanding these spectrospatial directivity patterns is central to bat echolocation.
Key concepts
- Far-field directivity
- This describes how focused or spread out a sound beam is when it travels far away from the source, like a flashlight beam. In this context, it measures the pattern of sound emitted by a bat and how that pattern changes depending on the direction of reception.
- Transfer function
- A transfer function mathematically describes how a specific part of an acoustic system, like an ear or noseleaf, modifies incoming sound waves. It tells researchers exactly what kind of frequency information is passed through or blocked by that anatomical structure.
- Posable models
- The dataset includes skeletons for the bat's pinnae (ears) and noseleaf that can be moved and rotated. This feature lets researchers test different head postures to see how changing the physical shape of the ears alters the resulting sound patterns.
Terminology
Summary
Echolocating bats shape their calls and the returning echoes with their noseleaf or mouth, head and pinnae Understanding these spectrospatial directivity patterns is central to bat echolocation. p. 0 Echolocating bats shape their calls and the returning echoes with their noseleaf or mouth,
head and pinnae. Understanding these spectrospatial directivity patterns is central to bat echolocation.Where the paper says it
Dataset Description
The paper describes a dataset of simulated far-field emission and reception directivities of seven echolocating bat species from three families (Phyllostomidae, Rhinolophidae and Vespertilionidae). For each species, the dataset contains several components including (1) a closed, simulation-ready head mesh derived from a micro-computed tomography (micro-CT) scan, (3) the complexvalued far-field transfer functions of both ears and of the emitter on the full sphere from 20 to 160 kHz, and (5) a skeleton for the pinnae and noseleaf, with which the meshes can be posed and simulated again. p. 0 In this Data Descriptor, we describe a dataset of
simulated emission and reception directivities of seven
species: Phyllostomus discolor, Micronycteris microtis, Macrophyllum macrophyllum, Rhinolophus rouxii,
Rhinolophus formosae, Eptesicus fuscus and Myotis
daubentonii (Table 1). For each species, it contains (1)
a closed, simulation-ready head mesh derived from a
micro-computed tomography (micro-CT) scan, (2) the
anatomical landmarks used as sources: the ear canal entrances, and the nostrils or the mouth, (3) the complexvalued far-field transfer functions of both ears and of
the emitter on the full sphere from 20 to 160 kHz, (4)
the transfer functions of up to four alternative positions of every ear and of the nostril pair, and (5) a
skeleton for the pinnae and the noseleaf, with which
the meshes can be posed and simulated again. Meshes,
landmarks and directivities share one coordinate frame,
so that they are spatially aligned without any further
registration. The emission patterns of the five nasal
emitters are BEM results, whereas the emitters of the
two mouth-emitting vespertilionids are analytical piston models calibrated to published beam widths, with
alternative radii, as their beam depends on the task
of the bat, and we keep this distinction throughout
the data and this paper.Where the paper says it
Where the paper says it
Modeling and Methodology
The acoustic model uses an acoustically rigid head with sound sources as small vibrating patches on its surface. The transfer function of an ear is computed by acoustic reciprocity, where a small patch at the ear canal entrance vibrates with uniform normal velocity. p. 3 Sources: ears and emitters
The transfer function of an ear is computed by acoustic reciprocity: a small patch at the ear canal entrance
(radius 0. 4 mm) vibrates with uniform normal velocity,
and the pressure it radiates to a far-field point, normalized as described below, equals the patch-averaged
pressure that a source with the same volume velocity
at that point produces at the ear. We therefore model the emission of these
two species with an analytical circular piston 28,29. The
piston approximates the open mouth and does not include the head. To place the sources of the other five species,
we derived a rule from these two: the best-matching
positions are those that radiate most strongly forward
(front/back ratio) and are most occluded by the surrounding anatomy, which is where pinna and noseleaf
focus sound forward. Each receiver is placed at the maximum of the standardized sum of the front/back ratio and the occlusion,
and the nostrils at the mirror-symmetric pair 1.Where the paper says it
Where the paper says it
Where the paper says it
Validation and Metrics
Technical validation involved numerical verification, comparing the solver with an analytical solution and independent data. Comparisons with earlier simulations showed consistency checks, such as the agreement between the dataset ERTFs and reference ERTFs of P. p. 3 Source placement
Where exactly the receiver sits at the ear canal entrance, and where the nostrils are, is a modelling choice
that the meshes do not fix, and our comparisons with
the earlier simulations showed that it matters (see Technical Validation). To quantify this, we computed dense
sets of candidate sources for the two species with earlier reference simulations, P. Coordinate frame and landmarks
All data are given in the frame of the original CT
meshes (for P. discolor in a rotated pose, see below),
with the axes renamed: X points forward, Y to the
bat’s left and Z up, in metres. 10
RMS di
fference (dB)
d GPU BEM vs NumCalc (M. microtis)
Left ear
Right ear
Emitter
Figure 3: Verification of the Solver. The relative
complex error ∥H − Href∥/∥Href∥ over all directions includes level and phase. The directivity index is the peak
power over the power averaged over the full sphere, and
the half-amplitude angle of a beam is the angular radius of a cone with the same solid angle as the region
within 6 dB of the peak.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Sensitivity Analysis
The dataset demonstrates sensitivity to source assumptions; for instance, alternative ear positions correlate with the nominal one with medians of 0.75–1.00. The directivity index for piston emitters is sensitive to the radius, changing by a median of 1. p. 8 00 (ears), 0. 83
(nostrils)
Piston radius Radius ±20 % Directivity index ±1. 5 kHz steps). Emitter: nasal (two vibrating nostril
patches, BEM) or oral (analytical piston of radius a, calibrated to the open-field search beam, with five alternative radii
from 2. daubentonii, measured HRTF 98th–99th percentile of 211 candidates
In-flight beam M. daubentonii Piston 15◦ narrower
Sensitivity
Source placement Alternative ear positions and nostril pairs Correlation 0. 70 for the median candidate, and 0. 81 when the rule is chosen without the held-out species.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Data Availability
The dataset is available at Zenodo 32 (https://doi.org/10.5281/zenodo. p. 4 Data Records
The dataset is available at Zenodo 32 (https://doi. 5281/zenodo. 5281/zenodo. 23729269), under a Creative
Commons Attribution 4. 23729269), under a Creative
Commons Attribution 4. 0 International (CC BY 4. 0 International (CC BY 4. 0)
licence. On Zenodo, every species folder is
one zip archive of the same name, the loader and the
examples are in code. zip, and the complete code that
produced the dataset is in bat ertf code-1. On Zenodo, every species folder is
one zip archive of the same name, the loader and the
examples are in code. zip, and the complete code that
produced the dataset is in bat ertf code-1. 0 International (CC BY 4. 0)
licence. txt lists the unpacked files. The total size of the dataset is 3.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Posable Models
Every mesh comes with a skeleton that allows for posing the pinnae and noseleaf, enabling studies of how posture shapes directivities. Posed meshes passed validity checks without self-intersections or stretched edges by more than a factor of two. p. 4 A posed mesh therefore has
the same triangles as the rest mesh, so that the source
patches and landmarks carry over. A validity check rejects poses that make the surface intersect itself, invert
faces or stretch edges by more than a factor of two. A validity check rejects poses that make the surface intersect itself, invert
faces or stretch edges by more than a factor of two. The
rest pose is the preserved posture of the specimen, for
which the transfer functions of the dataset were computed. The segmented surfaces
contain the head, including the pinnae with tragus and,
where present, the noseleaf, truncated behind the skull,
not containing body or wings such as in previous work 7
. The pinnae are in the posture in which the specimens
were preserved, and the mouths of all specimens are
closed (Supplementary Fig.Where the paper says it
Where the paper says it
Where the paper says it
Comparison with Measurements
The pipeline reproduces a measured HRTF as well as Mesh2HRTF does, showing that both simulations deviate from measurements in the same way. The alignment between measurement directions and the mesh requires an additional rotation of about 26◦ to increase median pattern correlations from 0. p. 8 Correlation: pattern correlation of two directivity patterns in dB (median
over frequency where a range of frequencies is compared). Relative error: relative complex error over all directions. 86 for the ears (means over the ten
placement frequencies), against 0. 94 for the best
possible position and 0. To place the sources of the other five species,
we derived a rule from these two: the best-matching
positions are those that radiate most strongly forward
(front/back ratio) and are most occluded by the surrounding anatomy, which is where pinna and noseleaf
focus sound forward. Each receiver is placed at the maximum of the standardized sum of the front/back ratio and the occlusion,
and the nostrils at the mirror-symmetric pair 1. daubentonii the analytical pistons calibrated to search calls in free flight in the
open field, the narrowest beam of these bats. In the laboratory, in confined spaces and when approaching prey, their beams
are considerably wider (see the alternative radii in the dataset and the Usage Notes). As every species is represented by a single specimen,
the dataset is a collection of specimen-based reference
models, not of species averages, and it does not capture
variation within a species. Furthermore, the responses
are far-field directivities of acoustically rigid heads, not
measurements and not full solutions of the near field. fuscus and M. p. 0 daubentonii and the in-flight emission beam of
that species of H¨afele et al. In this Data Descriptor, we describe a dataset of
simulated emission and reception directivities of seven
species: Phyllostomus discolor, Micronycteris microtis, Macrophyllum macrophyllum, Rhinolophus rouxii,
Rhinolophus formosae, Eptesicus fuscus and Myotis
daubentonii (Table 1). Meshes,
landmarks and directivities share one coordinate frame,
so that they are spatially aligned without any further
registration. The emission patterns of the five nasal
emitters are BEM results, whereas the emitters of the
two mouth-emitting vespertilionids are analytical piston models calibrated to published beam widths, with
alternative radii, as their beam depends on the task
of the bat, and we keep this distinction throughout
the data and this paper. Source placement
Where exactly the receiver sits at the ear canal entrance, and where the nostrils are, is a modelling choice
that the meshes do not fix, and our comparisons with
the earlier simulations showed that it matters (see Technical Validation). To quantify this, we computed dense
sets of candidate sources for the two species with earlier reference simulations, P. In this Data Descriptor, we describe a dataset of
simulated emission and reception directivities of seven
species: Phyllostomus discolor, Micronycteris microtis, Macrophyllum macrophyllum, Rhinolophus rouxii,
Rhinolophus formosae, Eptesicus fuscus and Myotis
daubentonii (Table 1). For each species, it contains (1)
a closed, simulation-ready head mesh derived from a
micro-computed tomography (micro-CT) scan, (2) the
anatomical landmarks used as sources: the ear canal entrances, and the nostrils or the mouth, (3) the complexvalued far-field transfer functions of both ears and of
the emitter on the full sphere from 20 to 160 kHz, (4)
the transfer functions of up to four alternative positions of every ear and of the nostril pair, and (5) a
skeleton for the pinnae and the noseleaf, with which
the meshes can be posed and simulated again.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Usage Notes
The HDF5 files can be read with any HDF5 library and complex values are stored as a compound of real and imaginary parts. Impulse responses follow from an inverse real FFT over frequency, after the band below 20 kHz and above the top of the band has been filled in with a smooth taper to zero. p. 4 With this referencing, the spherical
spreading and the propagation delay from the origin
are removed, and the phase of H is the phase relative
to a wave arriving at (or leaving from) the origin. An
inverse FFT over frequency therefore yields an impulse
response with t = 0 at the origin, which, for directions
on the side of the source, starts up to |⃗p|/c (about 60 µs)
before t = 0. Supplementary Video 1 (also available at https://youtu. be/
Q9Kxj9oIEwg) shows the seven heads turning with the
beam patterns of their left ear and emitter, frequency
sweeps of the left ear, the emitter and the two-way
ERTF of all species from 20 to 160 kHz, and the posable models of M. The magnitudes at 1 m represent the farfield directivity (pattern correlations of at least 0. 992
with the response at 20 m), while the phase still contains the curvature of the wavefront over the head (relative complex differences of up to 0. For the other
species, it follows the orientation of the scan, so that
the pitch of the head differs between species (Fig. We therefore did not rotate the data into a common
anatomical frame, which would require choices (for instance, of a skull-based horizontal plane) that are better
left to the user. 5 to 5 mm from the nominal
ones, which show how strongly a result depends on this
choice. The mouth landmark of the oral emitters was
placed by hand on the midline at the lip seam. fuscus and M. daubentonii the analytical pistons calibrated to search calls in free flight in the
open field, the narrowest beam of these bats. The piston contains no head, so that
its rear hemisphere, and the two-way ERTF built from
it, cannot be compared with those of the BEM emitters. We calibrated the radius to published field measurements of the beam width: 6. daubentonii the analytical pistons calibrated to search calls in free flight in the
open field, the narrowest beam of these bats. In the laboratory, in confined spaces and when approaching prey, their beams
are considerably wider (see the alternative radii in the dataset and the Usage Notes). 5 mm in
steps of 1 mm, which span the published beams of these
species from the laboratory (3. 7 mm28,31) to the
open field, and the released code computes the piston
for any radius. json Species, emitter type, landmarks, alternative source positions, piston parameters,
frame rotations, mesh statistics and conventions. rig/ Posable model: skeleton (rig. On the full
sphere, at 80, 120 and 160 kHz, the relative complex difference between the two solvers on this mesh is at most
0. As both use the same discretization, this shows
that our implementation solves the same discrete problem as NumCalc, but it does not test the discretization
error, which the next section addresses. 03 for the emitter, but a new
remesh of the CT surface at half the edge length differs
by 0. At high frequencies,
the uncertainty of the data is therefore dominated by
how faithfully the mesh represents the fine anatomy,
rather than by the number of elements per wavelength. Every transfer function is referenced to the free-field pressure at the same point of a monopole at the origin with
the same volume velocity Q as the source:
H(ψ, f) =
p(R ⃗u(ψ))
pf f (R)
∗
, pf f (R) = −jωρQ·
e
jkR
4πR
,
(3)
as in Mesh2HRTF21, where the complex conjugate converts the result into the e
+jωt convention of the discrete
Fourier transform. With this referencing, the spherical
spreading and the propagation delay from the origin
are removed, and the phase of H is the phase relative
to a wave arriving at (or leaving from) the origin.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI systems can simulate bat biosonar by using a dataset of simulated far-field emission and reception directivities for seven species, allowing for modeling of
echo-based localization
in complex environments. -
The system can perform comparative studies of how morphology shapes directivity by processing the provided data on head meshes and their corresponding transfer functions, enabling researchers to see
how morphology shapes directivity.
-
AI systems can be trained as training data for acoustic learning by utilizing the full dataset to study
acoustic learning,
which is central to bat echolocation. -
The system can be used for the design of biomimetic sonar by leveraging the analytical piston models calibrated to published beam widths, allowing users to compute
the piston for any radius.
-
AI systems can perform head-movement models by utilizing the
posable head models
and their associated skeletons, which allows for checking that posed meshesremain valid inputs for the BEM.
Abstract
Echolocating bats shape their calls and the returning echoes with their noseleaf or mouth, head and pinnae. Understanding these spectrospatial directivity patterns is central to bat echolocation. We describe a dataset of simulated far-field emission and reception directivities of seven echolocating bat species from three families (Phyllostomidae, Rhinolophidae and Vespertilionidae). For each species, the head of one preserved specimen was scanned with micro-computed tomography and converted into a closed surface mesh, on which the responses of both ears and of the emitter were computed with a boundary element method or, for the two mouth-emitting species, with an analytical piston model. The complex-valued responses cover the full sphere at 2.5 deg resolution from 20 to 160 kHz in 0.5 kHz steps. They come with the meshes, the landmarks of ears, nostrils and mouth, alternative source positions and skeletons for posing the pinnae and noseleaves, all in one frame, and with Python and MATLAB code. The technical validation covers solver verification, mesh convergence, comparisons with earlier simulations and measurements, and sensitivity analyses.
Related papers
- WavCube: Unifying Speech Representation for Understanding and Generation via Semantic-Acoustic Joint Modeling
- From perception to production: how acoustic invariance facilitates articulatory learning in a self-supervised vocal imitation model
- Read What You Hear: Reference-Free Hypotheses Evaluation with Acoustic Discrepancy
- pyFDN: A Python Toolbox for Feedback Delay Network Design, Analysis, and Optimization
- GigaSpeechBench: A Real-World Multilingual Speech-to-Text Benchmark
- DM-ASR: Diarization-aware Multi-speaker ASR with Unified Offline and Online Decoding with Large Language Models