mudler commited on
Commit
0602016
·
verified ·
1 Parent(s): bf0af9f

Add Nemotron-3-Diarization (speaker diarization) F16 and Q8_0

Browse files

Converted with parakeet.cpp 238057c. Segments match NeMo; see docs/diarization.md.

.gitattributes CHANGED
@@ -88,3 +88,5 @@ nemotron-3.5-asr-streaming-0.6b-q8_0.gguf filter=lfs diff=lfs merge=lfs -text
88
  nemotron-3.5-asr-streaming-0.6b-q6_k.gguf filter=lfs diff=lfs merge=lfs -text
89
  nemotron-3.5-asr-streaming-0.6b-q5_k.gguf filter=lfs diff=lfs merge=lfs -text
90
  nemotron-3.5-asr-streaming-0.6b-q4_k.gguf filter=lfs diff=lfs merge=lfs -text
 
 
 
88
  nemotron-3.5-asr-streaming-0.6b-q6_k.gguf filter=lfs diff=lfs merge=lfs -text
89
  nemotron-3.5-asr-streaming-0.6b-q5_k.gguf filter=lfs diff=lfs merge=lfs -text
90
  nemotron-3.5-asr-streaming-0.6b-q4_k.gguf filter=lfs diff=lfs merge=lfs -text
91
+ nemotron-3-diarization-f16.gguf filter=lfs diff=lfs merge=lfs -text
92
+ nemotron-3-diarization-q8_0.gguf filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -9,6 +9,7 @@ tags:
9
  - ggml
10
  - cpp-inference
11
  - nemo
 
12
  pipeline_tag: automatic-speech-recognition
13
  base_model:
14
  - nvidia/parakeet-tdt_ctc-110m
@@ -21,6 +22,7 @@ base_model:
21
  - nvidia/parakeet-rnnt-1.1b
22
  - nvidia/parakeet-tdt-1.1b
23
  - nvidia/parakeet-tdt_ctc-1.1b
 
24
  ---
25
 
26
  # Parakeet GGUF — models for parakeet.cpp
@@ -153,6 +155,21 @@ Source: [nvidia/parakeet-tdt_ctc-1.1b](https://huggingface.co/nvidia/parakeet-td
153
 
154
  > WER (word error rate) is computed against the upstream NeMo reference on `tests/fixtures/speech.wav` (LibriSpeech `2086-149220-0033`, ~7.4 s, English). 0.0 = byte-for-byte identical transcript. See [parity.md](https://github.com/mudler/parakeet.cpp/blob/main/docs/parity.md) and [quantization.md](https://github.com/mudler/parakeet.cpp/blob/main/docs/quantization.md).
155
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
156
  ## Quantization notes
157
 
158
  Quantization is applied **only** to the large linear weights fed directly into `ggml_mul_mat` (encoder FFN + attention projections, subsampling output projection, joint enc/pred projections). All other tensors (mel filterbank, LSTM prediction net, conv kernels, batch_norm stats, norms, biases, embeddings) stay F32.
@@ -176,4 +193,4 @@ build/examples/cli/parakeet-cli transcribe \
176
 
177
  ## License
178
 
179
- The GGUF weights are derived from the NVIDIA NeMo Parakeet checkpoints, released under the [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) license. The parakeet.cpp runtime is MIT-licensed.
 
9
  - ggml
10
  - cpp-inference
11
  - nemo
12
+ - speaker-diarization
13
  pipeline_tag: automatic-speech-recognition
14
  base_model:
15
  - nvidia/parakeet-tdt_ctc-110m
 
22
  - nvidia/parakeet-rnnt-1.1b
23
  - nvidia/parakeet-tdt-1.1b
24
  - nvidia/parakeet-tdt_ctc-1.1b
25
+ - nvidia/Nemotron-3-Diarization
26
  ---
27
 
28
  # Parakeet GGUF — models for parakeet.cpp
 
155
 
156
  > WER (word error rate) is computed against the upstream NeMo reference on `tests/fixtures/speech.wav` (LibriSpeech `2086-149220-0033`, ~7.4 s, English). 0.0 = byte-for-byte identical transcript. See [parity.md](https://github.com/mudler/parakeet.cpp/blob/main/docs/parity.md) and [quantization.md](https://github.com/mudler/parakeet.cpp/blob/main/docs/quantization.md).
157
 
158
+ ### nemotron-3-diarization
159
+
160
+ Source: [nvidia/Nemotron-3-Diarization](https://huggingface.co/nvidia/Nemotron-3-Diarization) · Speaker diarization (Sortformer, up to 8 speakers, streaming speaker cache) · License: [OpenMDW 1.1](https://openmdw.ai/license/1-1/)
161
+
162
+ | File | Variant | Size | Segments vs NeMo |
163
+ |---|---|---:|---:|
164
+ | `nemotron-3-diarization-f16.gguf` ← **recommended** | F16 | 191.4 MB | identical |
165
+ | `nemotron-3-diarization-q8_0.gguf` | Q8_0 | 103.6 MB | identical |
166
+
167
+ > Checked against NeMo on a 23.6 s and a 68.5 s two-speaker clip (offline and streaming: every segment identical, to the 10 ms frame) and a 12.3 min three-speaker clip (F16 100%, Q8_0 99.9% of speech frames). Answers "who spoke when"; pair it with any ASR model above for speaker-attributed transcripts. See [diarization.md](https://github.com/mudler/parakeet.cpp/blob/master/docs/diarization.md).
168
+
169
+ ```bash
170
+ build/examples/cli/diarize models/nemotron-3-diarization-f16.gguf meeting.wav
171
+ ```
172
+
173
  ## Quantization notes
174
 
175
  Quantization is applied **only** to the large linear weights fed directly into `ggml_mul_mat` (encoder FFN + attention projections, subsampling output projection, joint enc/pred projections). All other tensors (mel filterbank, LSTM prediction net, conv kernels, batch_norm stats, norms, biases, embeddings) stay F32.
 
193
 
194
  ## License
195
 
196
+ The Parakeet GGUF weights are derived from the NVIDIA NeMo Parakeet checkpoints, released under the [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) license. `nemotron-3-diarization-*.gguf` is derived from nvidia/Nemotron-3-Diarization and is governed by the [OpenMDW License Agreement, version 1.1](https://openmdw.ai/license/1-1/). The parakeet.cpp runtime is MIT-licensed.
nemotron-3-diarization-f16.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3dca41a92162af6edf96276c4d2c397b034b55abf9eb2662a58cc7bc8c872683
3
+ size 200715584
nemotron-3-diarization-q8_0.gguf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:76c5bb1fb20d82706142ad32769b7ab496d2458489473a000fd7074c52ceec22
3
+ size 108674624