murmur-asr 0.2.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,202 @@
1
+
2
+ Apache License
3
+ Version 2.0, January 2004
4
+ http://www.apache.org/licenses/
5
+
6
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
7
+
8
+ 1. Definitions.
9
+
10
+ "License" shall mean the terms and conditions for use, reproduction,
11
+ and distribution as defined by Sections 1 through 9 of this document.
12
+
13
+ "Licensor" shall mean the copyright owner or entity authorized by
14
+ the copyright owner that is granting the License.
15
+
16
+ "Legal Entity" shall mean the union of the acting entity and all
17
+ other entities that control, are controlled by, or are under common
18
+ control with that entity. For the purposes of this definition,
19
+ "control" means (i) the power, direct or indirect, to cause the
20
+ direction or management of such entity, whether by contract or
21
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
22
+ outstanding shares, or (iii) beneficial ownership of such entity.
23
+
24
+ "You" (or "Your") shall mean an individual or Legal Entity
25
+ exercising permissions granted by this License.
26
+
27
+ "Source" form shall mean the preferred form for making modifications,
28
+ including but not limited to software source code, documentation
29
+ source, and configuration files.
30
+
31
+ "Object" form shall mean any form resulting from mechanical
32
+ transformation or translation of a Source form, including but
33
+ not limited to compiled object code, generated documentation,
34
+ and conversions to other media types.
35
+
36
+ "Work" shall mean the work of authorship, whether in Source or
37
+ Object form, made available under the License, as indicated by a
38
+ copyright notice that is included in or attached to the work
39
+ (an example is provided in the Appendix below).
40
+
41
+ "Derivative Works" shall mean any work, whether in Source or Object
42
+ form, that is based on (or derived from) the Work and for which the
43
+ editorial revisions, annotations, elaborations, or other modifications
44
+ represent, as a whole, an original work of authorship. For the purposes
45
+ of this License, Derivative Works shall not include works that remain
46
+ separable from, or merely link (or bind by name) to the interfaces of,
47
+ the Work and Derivative Works thereof.
48
+
49
+ "Contribution" shall mean any work of authorship, including
50
+ the original version of the Work and any modifications or additions
51
+ to that Work or Derivative Works thereof, that is intentionally
52
+ submitted to Licensor for inclusion in the Work by the copyright owner
53
+ or by an individual or Legal Entity authorized to submit on behalf of
54
+ the copyright owner. For the purposes of this definition, "submitted"
55
+ means any form of electronic, verbal, or written communication sent
56
+ to the Licensor or its representatives, including but not limited to
57
+ communication on electronic mailing lists, source code control systems,
58
+ and issue tracking systems that are managed by, or on behalf of, the
59
+ Licensor for the purpose of discussing and improving the Work, but
60
+ excluding communication that is conspicuously marked or otherwise
61
+ designated in writing by the copyright owner as "Not a Contribution."
62
+
63
+ "Contributor" shall mean Licensor and any individual or Legal Entity
64
+ on behalf of whom a Contribution has been received by Licensor and
65
+ subsequently incorporated within the Work.
66
+
67
+ 2. Grant of Copyright License. Subject to the terms and conditions of
68
+ this License, each Contributor hereby grants to You a perpetual,
69
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
70
+ copyright license to reproduce, prepare Derivative Works of,
71
+ publicly display, publicly perform, sublicense, and distribute the
72
+ Work and such Derivative Works in Source or Object form.
73
+
74
+ 3. Grant of Patent License. Subject to the terms and conditions of
75
+ this License, each Contributor hereby grants to You a perpetual,
76
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
77
+ (except as stated in this section) patent license to make, have made,
78
+ use, offer to sell, sell, import, and otherwise transfer the Work,
79
+ where such license applies only to those patent claims licensable
80
+ by such Contributor that are necessarily infringed by their
81
+ Contribution(s) alone or by combination of their Contribution(s)
82
+ with the Work to which such Contribution(s) was submitted. If You
83
+ institute patent litigation against any entity (including a
84
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
85
+ or a Contribution incorporated within the Work constitutes direct
86
+ or contributory patent infringement, then any patent licenses
87
+ granted to You under this License for that Work shall terminate
88
+ as of the date such litigation is filed.
89
+
90
+ 4. Redistribution. You may reproduce and distribute copies of the
91
+ Work or Derivative Works thereof in any medium, with or without
92
+ modifications, and in Source or Object form, provided that You
93
+ meet the following conditions:
94
+
95
+ (a) You must give any other recipients of the Work or
96
+ Derivative Works a copy of this License; and
97
+
98
+ (b) You must cause any modified files to carry prominent notices
99
+ stating that You changed the files; and
100
+
101
+ (c) You must retain, in the Source form of any Derivative Works
102
+ that You distribute, all copyright, patent, trademark, and
103
+ attribution notices from the Source form of the Work,
104
+ excluding those notices that do not pertain to any part of
105
+ the Derivative Works; and
106
+
107
+ (d) If the Work includes a "NOTICE" text file as part of its
108
+ distribution, then any Derivative Works that You distribute must
109
+ include a readable copy of the attribution notices contained
110
+ within such NOTICE file, excluding those notices that do not
111
+ pertain to any part of the Derivative Works, in at least one
112
+ of the following places: within a NOTICE text file distributed
113
+ as part of the Derivative Works; within the Source form or
114
+ documentation, if provided along with the Derivative Works; or,
115
+ within a display generated by the Derivative Works, if and
116
+ wherever such third-party notices normally appear. The contents
117
+ of the NOTICE file are for informational purposes only and
118
+ do not modify the License. You may add Your own attribution
119
+ notices within Derivative Works that You distribute, alongside
120
+ or as an addendum to the NOTICE text from the Work, provided
121
+ that such additional attribution notices cannot be construed
122
+ as modifying the License.
123
+
124
+ You may add Your own copyright statement to Your modifications and
125
+ may provide additional or different license terms and conditions
126
+ for use, reproduction, or distribution of Your modifications, or
127
+ for any such Derivative Works as a whole, provided Your use,
128
+ reproduction, and distribution of the Work otherwise complies with
129
+ the conditions stated in this License.
130
+
131
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
132
+ any Contribution intentionally submitted for inclusion in the Work
133
+ by You to the Licensor shall be under the terms and conditions of
134
+ this License, without any additional terms or conditions.
135
+ Notwithstanding the above, nothing herein shall supersede or modify
136
+ the terms of any separate license agreement you may have executed
137
+ with Licensor regarding such Contributions.
138
+
139
+ 6. Trademarks. This License does not grant permission to use the trade
140
+ names, trademarks, service marks, or product names of the Licensor,
141
+ except as required for reasonable and customary use in describing the
142
+ origin of the Work and reproducing the content of the NOTICE file.
143
+
144
+ 7. Disclaimer of Warranty. Unless required by applicable law or
145
+ agreed to in writing, Licensor provides the Work (and each
146
+ Contributor provides its Contributions) on an "AS IS" BASIS,
147
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
148
+ implied, including, without limitation, any warranties or conditions
149
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
150
+ PARTICULAR PURPOSE. You are solely responsible for determining the
151
+ appropriateness of using or redistributing the Work and assume any
152
+ risks associated with Your exercise of permissions under this License.
153
+
154
+ 8. Limitation of Liability. In no event and under no legal theory,
155
+ whether in tort (including negligence), contract, or otherwise,
156
+ unless required by applicable law (such as deliberate and grossly
157
+ negligent acts) or agreed to in writing, shall any Contributor be
158
+ liable to You for damages, including any direct, indirect, special,
159
+ incidental, or consequential damages of any character arising as a
160
+ result of this License or out of the use or inability to use the
161
+ Work (including but not limited to damages for loss of goodwill,
162
+ work stoppage, computer failure or malfunction, or any and all
163
+ other commercial damages or losses), even if such Contributor
164
+ has been advised of the possibility of such damages.
165
+
166
+ 9. Accepting Warranty or Additional Liability. While redistributing
167
+ the Work or Derivative Works thereof, You may choose to offer,
168
+ and charge a fee for, acceptance of support, warranty, indemnity,
169
+ or other liability obligations and/or rights consistent with this
170
+ License. However, in accepting such obligations, You may act only
171
+ on Your own behalf and on Your sole responsibility, not on behalf
172
+ of any other Contributor, and only if You agree to indemnify,
173
+ defend, and hold each Contributor harmless for any liability
174
+ incurred by, or claims asserted against, such Contributor by reason
175
+ of your accepting any such warranty or additional liability.
176
+
177
+ END OF TERMS AND CONDITIONS
178
+
179
+ APPENDIX: How to apply the Apache License to your work.
180
+
181
+ To apply the Apache License to your work, attach the following
182
+ boilerplate notice, with the fields enclosed by brackets "[]"
183
+ replaced with your own identifying information. (Don't include
184
+ the brackets!) The text should be enclosed in the appropriate
185
+ comment syntax for the file format. We also recommend that a
186
+ file or class name and description of purpose be included on the
187
+ same "printed page" as the copyright notice for easier
188
+ identification within third-party archives.
189
+
190
+ Copyright [yyyy] [name of copyright owner]
191
+
192
+ Licensed under the Apache License, Version 2.0 (the "License");
193
+ you may not use this file except in compliance with the License.
194
+ You may obtain a copy of the License at
195
+
196
+ http://www.apache.org/licenses/LICENSE-2.0
197
+
198
+ Unless required by applicable law or agreed to in writing, software
199
+ distributed under the License is distributed on an "AS IS" BASIS,
200
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
201
+ See the License for the specific language governing permissions and
202
+ limitations under the License.
@@ -0,0 +1,5 @@
1
+ Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking
2
+ Copyright 2026 uw-syfi
3
+
4
+ Developed by Wei-Tzu Lee, Keisuke Kamahori and Baris Kasikci.
5
+ If you use Murmur in your work, please cite https://arxiv.org/abs/2606.01483
@@ -0,0 +1,197 @@
1
+ Metadata-Version: 2.4
2
+ Name: murmur-asr
3
+ Version: 0.2.0
4
+ Summary: Efficient long-context ASR inference with hierarchical chunking (target chunks + draft windows on vLLM).
5
+ Author: uw-syfi
6
+ License-Expression: Apache-2.0
7
+ Project-URL: Homepage, https://github.com/uw-syfi/Murmur
8
+ Project-URL: Repository, https://github.com/uw-syfi/Murmur
9
+ Project-URL: Issues, https://github.com/uw-syfi/Murmur/issues
10
+ Project-URL: Paper, https://arxiv.org/abs/2606.01483
11
+ Keywords: asr,speech-recognition,vllm,speculative-decoding,diarization
12
+ Classifier: Development Status :: 4 - Beta
13
+ Classifier: Intended Audience :: Science/Research
14
+ Classifier: Programming Language :: Python :: 3
15
+ Classifier: Programming Language :: Python :: 3.10
16
+ Classifier: Topic :: Multimedia :: Sound/Audio :: Speech
17
+ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
18
+ Requires-Python: >=3.10
19
+ Description-Content-Type: text/markdown
20
+ License-File: LICENSE
21
+ License-File: NOTICE
22
+ Requires-Dist: numpy
23
+ Requires-Dist: requests
24
+ Requires-Dist: scipy
25
+ Requires-Dist: soundfile
26
+ Requires-Dist: torch
27
+ Requires-Dist: torchaudio
28
+ Provides-Extra: diarize
29
+ Requires-Dist: pyannote.audio>=4.0; extra == "diarize"
30
+ Provides-Extra: drafts
31
+ Requires-Dist: faster-whisper>=1.0; extra == "drafts"
32
+ Requires-Dist: opencc; extra == "drafts"
33
+ Provides-Extra: eval
34
+ Requires-Dist: jiwer; extra == "eval"
35
+ Requires-Dist: meeteval; extra == "eval"
36
+ Requires-Dist: pyannote.metrics; extra == "eval"
37
+ Provides-Extra: all
38
+ Requires-Dist: murmur-asr[diarize,drafts,eval]; extra == "all"
39
+ Provides-Extra: server
40
+ Requires-Dist: transformers; extra == "server"
41
+ Provides-Extra: test
42
+ Requires-Dist: pytest; extra == "test"
43
+ Requires-Dist: jiwer; extra == "test"
44
+ Dynamic: license-file
45
+
46
+ # Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking
47
+
48
+ [![arXiv](https://img.shields.io/badge/arXiv-2606.01483-b31b1b.svg)](https://arxiv.org/abs/2606.01483)
49
+
50
+ **News**
51
+ - 2026/11/13: [Blog post](https://syfi.cs.washington.edu/blog/2026-11-13-murmur/) on Murmur is out.
52
+ - 2026/11/04: [Preprint v2](https://arxiv.org/abs/2606.01483v2) is out.
53
+ - 2026/06/01: [Preprint](https://arxiv.org/abs/2606.01483v1) is released on arXiv.
54
+
55
+ Murmur is an inference system that serves long-context ASR models, such as **VibeVoice-ASR** and
56
+ **MOSS-Transcribe-Diarize**, on long audio. These models transcribe an hour-long recording in a single
57
+ pass, which is slow and can fail entirely on a repetition loop.
58
+
59
+ - **Across chunks:** Murmur decodes ~5-minute target chunks in parallel. This intermediate size gives
60
+ the model enough context to match single pass in most cases, and a loop costs at most one chunk.
61
+ - **Within each chunk:** hierarchical chunking splits each chunk into 30 s draft windows. A small
62
+ Whisper model transcribes them, and structure-aware drafting adds the timestamps and speaker tags,
63
+ giving drafts for speculative decoding. The long-context model verifies every token, so decoding is
64
+ faster and the output is unchanged.
65
+
66
+ Overall, Murmur is **2.5x faster** than single pass on average (2.0-2.9x), end to end including
67
+ diarization, and lowers tcpWER by 2.0 points on average.
68
+
69
+ ![Hierarchical chunking](https://raw.githubusercontent.com/uw-syfi/Murmur/main/docs/figures/hierarchy.png)
70
+
71
+ ## Demo
72
+
73
+ Single pass, Murmur and WhisperX transcribing the same 32-minute AMI meeting on one H100, replayed at 8x.
74
+
75
+ https://github.com/user-attachments/assets/e2897c2a-cc33-4021-8e2d-7dfbac46f8c9
76
+
77
+ ## Results
78
+
79
+ Murmur against each model's own single pass on four meeting datasets. Each cell is tcpWER / latency
80
+ per recording (lower is better), on one H100 with vLLM, with latency including diarization.
81
+
82
+ | Model | System | AMI-IHM | AMI-SDM | AliMeeting | AISHELL-4 |
83
+ |---|---|---|---|---|---|
84
+ | VibeVoice-ASR | single pass | 30.1% / 109 s | 39.6% / 100 s | 41.7% / 123 s | **30.3%** / 111 s |
85
+ | | **Murmur** | **24.5% / 38 s** | **32.9% / 39 s** | **38.7% / 47 s** | 30.5% / **46 s** |
86
+ | MOSS-Transcribe-Diarize | single pass | 17.5% / 54 s | 29.6% / 77 s | **13.6%** / 91 s | **14.1%** / 51 s |
87
+ | | **Murmur** | **17.3% / 26 s** | **22.8% / 26 s** | 17.5% / **32 s** | 16.6% / **26 s** |
88
+
89
+ <sub>Bold: better of single pass and Murmur. WER, cpWER and DER are in the paper. AMI scores use a
90
+ corrected reference reader (2-6 points lower than earlier arXiv versions).</sub>
91
+
92
+ ## Supported models
93
+
94
+ Murmur runs each model on its own vLLM server and talks to it through a small client class.
95
+
96
+ | Model | `Murmur.load(...)` | Draft windows | vLLM server |
97
+ |---|---|---|---|
98
+ | [VibeVoice-ASR](https://huggingface.co/microsoft/VibeVoice-ASR) (8.5B) | `"vibevoice"` | on | vLLM 0.14 + VibeVoice plugin |
99
+ | [MOSS-Transcribe-Diarize](https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize) (0.9B) | `"moss"` | off | vLLM nightly |
100
+
101
+ Both are tested on English and Mandarin. To add a model, add a client class in `murmur/models/`.
102
+
103
+ ## Install
104
+
105
+ The `murmur` client and each model's vLLM server use separate environments, because the two models
106
+ need different vLLM builds.
107
+
108
+ ```bash
109
+ conda create -n murmur python=3.10 -y && conda activate murmur
110
+ pip install "murmur-asr[all]"
111
+ # or, from source: git clone https://github.com/uw-syfi/Murmur.git && pip install -e "Murmur[all]"
112
+ huggingface-cli login # accept the terms of pyannote/speaker-diarization-3.1 first
113
+ ```
114
+
115
+ <details>
116
+ <summary>vLLM servers (VibeVoice-ASR, MOSS-Transcribe-Diarize)</summary>
117
+
118
+ ```bash
119
+ # VibeVoice-ASR: vLLM 0.14 + the VibeVoice plugin
120
+ conda create -n vibevoice-vllm python=3.10 -y && conda activate vibevoice-vllm
121
+ pip install vllm==0.14.0
122
+ git clone https://github.com/microsoft/VibeVoice.git && pip install -e VibeVoice
123
+ pip install "murmur-asr[server]"
124
+
125
+ # MOSS-Transcribe-Diarize: vLLM nightly (tested: 0.23.1rc1.dev949)
126
+ conda create -n moss-vllm python=3.10 -y && conda activate moss-vllm
127
+ pip install -U vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
128
+ pip install "murmur-asr[server]"
129
+
130
+ # so the client can start the servers itself
131
+ export MURMUR_VLLM_VIBEVOICE=$CONDA_PREFIX/../vibevoice-vllm/bin/vllm
132
+ export MURMUR_VLLM_MOSS=$CONDA_PREFIX/../moss-vllm/bin/vllm
133
+ ```
134
+ </details>
135
+
136
+ ## Usage
137
+
138
+ `Murmur.load` starts the model's vLLM server, and `transcribe` returns segments with speakers and
139
+ timestamps.
140
+
141
+ ```python
142
+ from murmur import Murmur
143
+
144
+ with Murmur.load("vibevoice") as m: # or Murmur.load("moss", language="zh")
145
+ print(m.transcribe("meeting.wav")) # [ 12.40- 15.10] SPEAKER_01: ...
146
+ ```
147
+
148
+ <details>
149
+ <summary>Options, CLI and a separately started server</summary>
150
+
151
+ `Murmur.load` options: `chunk_s` (default 330), `batch_size` (default 8), `draft_windows`,
152
+ `language` (`"en"` or `"zh"`), `diarize`, `gpu`. `transcribe(..., single_pass=True)` runs the
153
+ single-pass baseline.
154
+
155
+ ```bash
156
+ # in the vibevoice-vllm environment
157
+ CUDA_VISIBLE_DEVICES=0 murmur serve vibevoice --port 8000 --draft-dir /tmp/murmur_drafts
158
+ # in the murmur environment
159
+ murmur transcribe meeting.wav --model vibevoice --url http://localhost:8000 --draft-dir /tmp/murmur_drafts
160
+ ```
161
+ Client and server must share the draft directory (same machine).
162
+ </details>
163
+
164
+ ## Benchmarks
165
+
166
+ `benchmarks/run.py` reproduces the paper's results, one model, dataset and system at a time.
167
+
168
+ ```bash
169
+ python benchmarks/run.py --model vibevoice --dataset ami_ihm --system murmur --mode latency
170
+ ```
171
+
172
+ <details>
173
+ <summary>Datasets, systems and scoring</summary>
174
+
175
+ ```bash
176
+ bash benchmarks/prepare/download_ami.sh && python benchmarks/prepare/build_ami_sessions.py # needs lhotse
177
+ python benchmarks/prepare/download_alimeeting_refs.py --archives Eval_Ali.tar.gz Test_Ali.tar.gz --raw_dir /tmp/ali # OpenSLR 119
178
+ python benchmarks/prepare/download_aishell4_refs.py --archive test.tar.gz --raw_dir /tmp/aishell4 # OpenSLR 111
179
+ ```
180
+
181
+ `--system`: `single`, `chunked` (greedy), `murmur` (as in the paper; `--draft-windows on|off`
182
+ overrides) or `ngram` (vLLM prompt lookup). `--mode latency` reports end-to-end latency per
183
+ recording; `--mode throughput` reports decoding time for a whole dataset. Scoring: WER (CER for
184
+ Mandarin) with jiwer, cpWER and tcpWER (5 s collar) with MeetEval, DER with pyannote.metrics.
185
+ </details>
186
+
187
+ ## Citation
188
+
189
+ ```bibtex
190
+ @article{murmur2026lee,
191
+ title={Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking},
192
+ author={Lee, Wei-Tzu and Kamahori, Keisuke and Kasikci, Baris},
193
+ journal={arXiv preprint arXiv:2606.01483},
194
+ year={2026},
195
+ url={https://arxiv.org/abs/2606.01483}
196
+ }
197
+ ```
@@ -0,0 +1,152 @@
1
+ # Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking
2
+
3
+ [![arXiv](https://img.shields.io/badge/arXiv-2606.01483-b31b1b.svg)](https://arxiv.org/abs/2606.01483)
4
+
5
+ **News**
6
+ - 2026/11/13: [Blog post](https://syfi.cs.washington.edu/blog/2026-11-13-murmur/) on Murmur is out.
7
+ - 2026/11/04: [Preprint v2](https://arxiv.org/abs/2606.01483v2) is out.
8
+ - 2026/06/01: [Preprint](https://arxiv.org/abs/2606.01483v1) is released on arXiv.
9
+
10
+ Murmur is an inference system that serves long-context ASR models, such as **VibeVoice-ASR** and
11
+ **MOSS-Transcribe-Diarize**, on long audio. These models transcribe an hour-long recording in a single
12
+ pass, which is slow and can fail entirely on a repetition loop.
13
+
14
+ - **Across chunks:** Murmur decodes ~5-minute target chunks in parallel. This intermediate size gives
15
+ the model enough context to match single pass in most cases, and a loop costs at most one chunk.
16
+ - **Within each chunk:** hierarchical chunking splits each chunk into 30 s draft windows. A small
17
+ Whisper model transcribes them, and structure-aware drafting adds the timestamps and speaker tags,
18
+ giving drafts for speculative decoding. The long-context model verifies every token, so decoding is
19
+ faster and the output is unchanged.
20
+
21
+ Overall, Murmur is **2.5x faster** than single pass on average (2.0-2.9x), end to end including
22
+ diarization, and lowers tcpWER by 2.0 points on average.
23
+
24
+ ![Hierarchical chunking](https://raw.githubusercontent.com/uw-syfi/Murmur/main/docs/figures/hierarchy.png)
25
+
26
+ ## Demo
27
+
28
+ Single pass, Murmur and WhisperX transcribing the same 32-minute AMI meeting on one H100, replayed at 8x.
29
+
30
+ https://github.com/user-attachments/assets/e2897c2a-cc33-4021-8e2d-7dfbac46f8c9
31
+
32
+ ## Results
33
+
34
+ Murmur against each model's own single pass on four meeting datasets. Each cell is tcpWER / latency
35
+ per recording (lower is better), on one H100 with vLLM, with latency including diarization.
36
+
37
+ | Model | System | AMI-IHM | AMI-SDM | AliMeeting | AISHELL-4 |
38
+ |---|---|---|---|---|---|
39
+ | VibeVoice-ASR | single pass | 30.1% / 109 s | 39.6% / 100 s | 41.7% / 123 s | **30.3%** / 111 s |
40
+ | | **Murmur** | **24.5% / 38 s** | **32.9% / 39 s** | **38.7% / 47 s** | 30.5% / **46 s** |
41
+ | MOSS-Transcribe-Diarize | single pass | 17.5% / 54 s | 29.6% / 77 s | **13.6%** / 91 s | **14.1%** / 51 s |
42
+ | | **Murmur** | **17.3% / 26 s** | **22.8% / 26 s** | 17.5% / **32 s** | 16.6% / **26 s** |
43
+
44
+ <sub>Bold: better of single pass and Murmur. WER, cpWER and DER are in the paper. AMI scores use a
45
+ corrected reference reader (2-6 points lower than earlier arXiv versions).</sub>
46
+
47
+ ## Supported models
48
+
49
+ Murmur runs each model on its own vLLM server and talks to it through a small client class.
50
+
51
+ | Model | `Murmur.load(...)` | Draft windows | vLLM server |
52
+ |---|---|---|---|
53
+ | [VibeVoice-ASR](https://huggingface.co/microsoft/VibeVoice-ASR) (8.5B) | `"vibevoice"` | on | vLLM 0.14 + VibeVoice plugin |
54
+ | [MOSS-Transcribe-Diarize](https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize) (0.9B) | `"moss"` | off | vLLM nightly |
55
+
56
+ Both are tested on English and Mandarin. To add a model, add a client class in `murmur/models/`.
57
+
58
+ ## Install
59
+
60
+ The `murmur` client and each model's vLLM server use separate environments, because the two models
61
+ need different vLLM builds.
62
+
63
+ ```bash
64
+ conda create -n murmur python=3.10 -y && conda activate murmur
65
+ pip install "murmur-asr[all]"
66
+ # or, from source: git clone https://github.com/uw-syfi/Murmur.git && pip install -e "Murmur[all]"
67
+ huggingface-cli login # accept the terms of pyannote/speaker-diarization-3.1 first
68
+ ```
69
+
70
+ <details>
71
+ <summary>vLLM servers (VibeVoice-ASR, MOSS-Transcribe-Diarize)</summary>
72
+
73
+ ```bash
74
+ # VibeVoice-ASR: vLLM 0.14 + the VibeVoice plugin
75
+ conda create -n vibevoice-vllm python=3.10 -y && conda activate vibevoice-vllm
76
+ pip install vllm==0.14.0
77
+ git clone https://github.com/microsoft/VibeVoice.git && pip install -e VibeVoice
78
+ pip install "murmur-asr[server]"
79
+
80
+ # MOSS-Transcribe-Diarize: vLLM nightly (tested: 0.23.1rc1.dev949)
81
+ conda create -n moss-vllm python=3.10 -y && conda activate moss-vllm
82
+ pip install -U vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
83
+ pip install "murmur-asr[server]"
84
+
85
+ # so the client can start the servers itself
86
+ export MURMUR_VLLM_VIBEVOICE=$CONDA_PREFIX/../vibevoice-vllm/bin/vllm
87
+ export MURMUR_VLLM_MOSS=$CONDA_PREFIX/../moss-vllm/bin/vllm
88
+ ```
89
+ </details>
90
+
91
+ ## Usage
92
+
93
+ `Murmur.load` starts the model's vLLM server, and `transcribe` returns segments with speakers and
94
+ timestamps.
95
+
96
+ ```python
97
+ from murmur import Murmur
98
+
99
+ with Murmur.load("vibevoice") as m: # or Murmur.load("moss", language="zh")
100
+ print(m.transcribe("meeting.wav")) # [ 12.40- 15.10] SPEAKER_01: ...
101
+ ```
102
+
103
+ <details>
104
+ <summary>Options, CLI and a separately started server</summary>
105
+
106
+ `Murmur.load` options: `chunk_s` (default 330), `batch_size` (default 8), `draft_windows`,
107
+ `language` (`"en"` or `"zh"`), `diarize`, `gpu`. `transcribe(..., single_pass=True)` runs the
108
+ single-pass baseline.
109
+
110
+ ```bash
111
+ # in the vibevoice-vllm environment
112
+ CUDA_VISIBLE_DEVICES=0 murmur serve vibevoice --port 8000 --draft-dir /tmp/murmur_drafts
113
+ # in the murmur environment
114
+ murmur transcribe meeting.wav --model vibevoice --url http://localhost:8000 --draft-dir /tmp/murmur_drafts
115
+ ```
116
+ Client and server must share the draft directory (same machine).
117
+ </details>
118
+
119
+ ## Benchmarks
120
+
121
+ `benchmarks/run.py` reproduces the paper's results, one model, dataset and system at a time.
122
+
123
+ ```bash
124
+ python benchmarks/run.py --model vibevoice --dataset ami_ihm --system murmur --mode latency
125
+ ```
126
+
127
+ <details>
128
+ <summary>Datasets, systems and scoring</summary>
129
+
130
+ ```bash
131
+ bash benchmarks/prepare/download_ami.sh && python benchmarks/prepare/build_ami_sessions.py # needs lhotse
132
+ python benchmarks/prepare/download_alimeeting_refs.py --archives Eval_Ali.tar.gz Test_Ali.tar.gz --raw_dir /tmp/ali # OpenSLR 119
133
+ python benchmarks/prepare/download_aishell4_refs.py --archive test.tar.gz --raw_dir /tmp/aishell4 # OpenSLR 111
134
+ ```
135
+
136
+ `--system`: `single`, `chunked` (greedy), `murmur` (as in the paper; `--draft-windows on|off`
137
+ overrides) or `ngram` (vLLM prompt lookup). `--mode latency` reports end-to-end latency per
138
+ recording; `--mode throughput` reports decoding time for a whole dataset. Scoring: WER (CER for
139
+ Mandarin) with jiwer, cpWER and tcpWER (5 s collar) with MeetEval, DER with pyannote.metrics.
140
+ </details>
141
+
142
+ ## Citation
143
+
144
+ ```bibtex
145
+ @article{murmur2026lee,
146
+ title={Murmur: Efficient Long-Context ASR Inference with Hierarchical Chunking},
147
+ author={Lee, Wei-Tzu and Kamahori, Keisuke and Kasikci, Baris},
148
+ journal={arXiv preprint arXiv:2606.01483},
149
+ year={2026},
150
+ url={https://arxiv.org/abs/2606.01483}
151
+ }
152
+ ```
@@ -0,0 +1,26 @@
1
+ """Murmur: efficient long-context ASR inference with hierarchical chunking.
2
+
3
+ from murmur import Murmur
4
+ m = Murmur.load("vibevoice")
5
+ print(m.transcribe("meeting.wav"))
6
+
7
+ Imports are lazy so that the vLLM server environment can import
8
+ ``murmur.drafting.vllm_proposer`` without the client's dependencies.
9
+ """
10
+
11
+ __version__ = "0.2.0"
12
+
13
+ __all__ = ["Murmur", "load", "Segment", "Transcript", "evaluate"]
14
+
15
+
16
+ def __getattr__(name):
17
+ if name in ("Murmur", "load"):
18
+ from . import api
19
+ return getattr(api, name)
20
+ if name in ("Segment", "Transcript"):
21
+ from . import types
22
+ return getattr(types, name)
23
+ if name == "evaluate":
24
+ from .metrics import evaluate
25
+ return evaluate
26
+ raise AttributeError(f"module 'murmur' has no attribute {name!r}")