README.md
48.0 KB · 566 lines · markdown Raw
1 ---
2 license: other
3 license_name: openmdw-1.1
4 license_link: https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b
5 base_model: nvidia/nemotron-3.5-asr-streaming-0.6b
6 base_model_relation: quantized
7 library_name: transcribe.cpp
8 pipeline_tag: automatic-speech-recognition
9 language:
10 - en
11 - es
12 - fr
13 - it
14 - pt
15 - nl
16 - de
17 - tr
18 - ru
19 - ar
20 - hi
21 - ja
22 - ko
23 - vi
24 - uk
25 - pl
26 - sv
27 - cs
28 - nb
29 - da
30 - bg
31 - fi
32 - hr
33 - sk
34 - zh
35 - hu
36 - ro
37 - et
38 tags:
39 - gguf
40 - transcribe.cpp
41 - asr
42 - speech-to-text
43 - parakeet
44 - conformer
45 - rnnt
46 - streaming
47 - cache-aware
48 - multilingual
49 transcribe_cpp:
50 wer_fleurs_en:
51 f32: 7.97
52 f16: 7.97
53 q8_0: 7.88
54 q6_k: 8.02
55 q5_k_m: 8.15
56 q4_k_m: 8.49
57 wer_librispeech_test_clean:
58 f32: 3.04
59 f16: 3.03
60 q8_0: 3.06
61 q6_k: 3.07
62 q5_k_m: 3.1
63 q4_k_m: 3.28
64 rtf_m4_max:
65 metal: 98
66 cpu: 29
67 rtf_ryzen_4750u:
68 vulkan: 14.5
69 cpu: 7.5
70 streaming: true
71 translate: false
72 lang_detect: true
73 timestamps: token
74 ---
75
76 # nemotron-3.5-asr-streaming-0.6b: transcribe.cpp GGUF
77
78 GGUF conversions of [nvidia/nemotron-3.5-asr-streaming-0.6b](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b) for use
79 with [transcribe.cpp](https://github.com/handy-computer/transcribe.cpp).
80
81 Ported from upstream commit
82 [24b151a](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b/commit/24b151a),
83 pinned 2026-06-08.
84 Validated against the NeMo reference at transcribe.cpp commit
85 [909e94e](https://github.com/handy-computer/transcribe.cpp/tree/909e94e)
86 on 2026-06-08.
87
88 Multilingual speech-to-text across 32 supported language-locales (the model's tokenizer recognizes 40, but 8 are adaptation-ready and need fine-tuning) with punctuation and capitalization. A 0.6B-parameter cache-aware streaming FastConformer encoder with a prompt-conditioned RNN-T transducer decoder; the target language is selected per call (--language en-US, fr-FR, de-DE, ...) and an auto mode emits a <lang-XX> tag. Ships both the offline path (att_context_size=[56, 13], 1.12s, headline accuracy) and runtime-selectable chunked streaming (--stream-chunk-ms 1120 --stream-att-right {0,3,6,13}).
89
90
91 ## Downloads
92
93 | Quantization | Download | Size | WER (FLEURS test en (en-US), offline att_context_size=[56, 13]) |
94 | --- | --- | ---: | ---: |
95 | F32 | [nemotron-3.5-asr-streaming-0.6b-F32.gguf](https://huggingface.co/handy-computer/nemotron-3.5-asr-streaming-0.6b-gguf/resolve/main/nemotron-3.5-asr-streaming-0.6b-F32.gguf) | 2.38 GB | 7.97% |
96 | F16 | [nemotron-3.5-asr-streaming-0.6b-F16.gguf](https://huggingface.co/handy-computer/nemotron-3.5-asr-streaming-0.6b-gguf/resolve/main/nemotron-3.5-asr-streaming-0.6b-F16.gguf) | 1.19 GB | 7.97% |
97 | Q8_0 | [nemotron-3.5-asr-streaming-0.6b-Q8_0.gguf](https://huggingface.co/handy-computer/nemotron-3.5-asr-streaming-0.6b-gguf/resolve/main/nemotron-3.5-asr-streaming-0.6b-Q8_0.gguf) | 716 MB | 7.88% |
98 | Q6_K | [nemotron-3.5-asr-streaming-0.6b-Q6_K.gguf](https://huggingface.co/handy-computer/nemotron-3.5-asr-streaming-0.6b-gguf/resolve/main/nemotron-3.5-asr-streaming-0.6b-Q6_K.gguf) | 593 MB | 8.02% |
99 | Q5_K_M | [nemotron-3.5-asr-streaming-0.6b-Q5_K_M.gguf](https://huggingface.co/handy-computer/nemotron-3.5-asr-streaming-0.6b-gguf/resolve/main/nemotron-3.5-asr-streaming-0.6b-Q5_K_M.gguf) | 534 MB | 8.15% |
100 | Q4_K_M | [nemotron-3.5-asr-streaming-0.6b-Q4_K_M.gguf](https://huggingface.co/handy-computer/nemotron-3.5-asr-streaming-0.6b-gguf/resolve/main/nemotron-3.5-asr-streaming-0.6b-Q4_K_M.gguf) | 473 MB | 8.49% |
101
102 WER measured on FLEURS test en (647 utterances), greedy RNN-T, --language en-US, whisper-normalizer scoring; the per-quant column is FLEURS en. NeMo reference baseline on the same manifest: 7.99% (NVIDIA self-reports 7.91% en-US). On LibriSpeech test-clean (2620 utterances) the same presets score F32 3.04 / F16 3.03 / Q8_0 3.06 / Q6_K 3.07 / Q5_K_M 3.10 / Q4_K_M 3.28, against a 3.03% NeMo reference.
103
104
105 ## Usage
106
107 Build transcribe.cpp from source:
108
109 ```bash
110 git clone git@github.com:handy-computer/transcribe.cpp.git
111 cd transcribe.cpp
112 cmake -B build && cmake --build build
113 ```
114
115 Run on a 16 kHz mono WAV:
116
117 ```bash
118 build/bin/transcribe-cli \
119 -m nemotron-3.5-asr-streaming-0.6b-Q8_0.gguf \
120 input.wav
121 ```
122
123 If your audio isn't already 16 kHz mono WAV, convert it first:
124
125 ```bash
126 ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav
127 ```
128
129 See the [transcribe.cpp model page](https://github.com/handy-computer/transcribe.cpp/blob/main/docs/models/nemotron-3.5-asr-streaming-0.6b.md) for performance
130 numbers, numerical validation, and reproduction steps.
131
132 ## License
133
134 Inherited from the base model: **OpenMDW-1.1**. See the
135 [upstream model card](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b) for full terms.
136
137 ---
138
139 ## Original Model Card
140
141 > The section below is reproduced from
142 > [nvidia/nemotron-3.5-asr-streaming-0.6b](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b) at commit
143 > `24b151a` for offline reference. The upstream card is the
144 > authoritative source.
145
146 # Nemotron 3.5 ASR
147
148 <style>
149 h1, h2, h3, h4, h5, h6 {
150 color: #76b900; /* NVIDIA green */
151 font-weight: 700;
152 }
153
154 hr {
155 border: none;
156 border-top: 1px solid #e5e7eb;
157 margin: 2rem 0;
158 }
159
160 /* Improve list spacing */
161 ul, ol {
162 margin-top: 0.5rem;
163 margin-bottom: 0.5rem;
164 }
165
166 /* Badge alignment consistency */
167 img {
168 display: inline;
169 vertical-align: middle;
170 }
171 </style>
172
173 <p align="center">
174 <a href="#model-architecture"><img src="https://img.shields.io/badge/Model_Arch-FastConformer--CacheAware--RNNT-lightgrey#model-badge" alt="Model architecture"/></a>
175 &nbsp;
176 <a href="#model-architecture"><img src="https://img.shields.io/badge/Params-600M-lightgrey#model-badge" alt="Model size"/></a>
177 &nbsp;
178 <a href="#supported-languages"><img src="https://img.shields.io/badge/Language-Multilingual-lightgrey#model-badge" alt="Language"/></a>
179 </p>
180
181 <p align="center">
182 <img src="model_overview.png" alt="Nemotron 3.5 ASR overview: multilingual audio across 40 language-locales is transcribed by a cache-aware FastConformer-RNNT model with language-ID prompting into punctuated text with an automatic language tag" width="900"/>
183 </p>
184
185 > [!Note]
186 > This model is the multilingual extension of [nvidia/nemotron-speech-streaming-en-0.6b](https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b), adding language-ID prompt conditioning to support transcription across **40 language-locales** from a single model.
187
188 **Nemotron 3.5 ASR** is a multilingual, streaming Automatic Speech Recognition (ASR) model engineered to deliver high-quality multilingual transcription across both low-latency streaming and high-throughput batch workloads. Developed by NVIDIA, this 600M parameter model transcribes speech into text with native support for punctuation and capitalization, and offers runtime flexibility with configurable chunk sizes, including 80ms, 160ms, 320ms, 560ms, and 1120ms.
189
190 By leveraging a state-of-the-art **Cache-Aware FastConformer-RNNT** architecture, the model eliminates redundant overlapping computations common in traditional "buffered" streaming. This allows it to process only new audio chunks while reusing cached encoder context, significantly improving computational efficiency and minimizing end-to-end delay without sacrificing accuracy.
191
192 It was trained on a massive ASR dataset and is engineered to perform across diverse and challenging acoustic conditions.
193
194 This model is ready for commercial use.
195
196 ---
197
198 ## License/Terms of Use
199
200 Governing Terms: Use of the model is governed by the [OpenMDW-1.1](https://openmdw.ai/license/1-1/) license.
201
202 ## Deployment Geography
203
204 Global
205
206 ## Use Case
207
208 This model is for transcription of multilingual audio.
209
210 ## Release Date
211
212 - Hugging Face [06/04/2026] via https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b
213
214 ## References
215
216 <a id="ref-1"></a>[1] [Stateful Conformer with Cache-based Inference for Streaming Automatic Speech Recognition](https://arxiv.org/abs/2312.17279)
217
218 <a id="ref-2"></a>[2] [Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition](https://arxiv.org/abs/2305.05084)
219
220 <a id="ref-3"></a>[3] [NVIDIA Granary](https://huggingface.co/datasets/nvidia/Granary)
221
222 <a id="ref-4"></a>[4] [NVIDIA NeMo Framework](https://github.com/NVIDIA/NeMo)
223
224 ## Why Choose Nemotron 3.5 ASR?
225
226 - 🌍 **Single Multilingual Model:** Transcribes 40 language-locales from one model through language-ID prompt conditioning, with optional automatic language detection.
227 - ⚡ **Native Streaming Architecture:** Cache-aware design enables efficient processing of continuous audio streams, designed and optimized for low-latency voice agent applications.
228 - 💰 **Improved Operational Efficiency:** Delivers superior throughput compared to traditional buffered streaming approaches. This allows for a higher number of parallel streams within the same GPU memory constraints, directly reducing operational costs for production environments.
229 - 🎛️ **Dynamic Runtime Flexibility:** Choose the optimal operating point on the latency-accuracy Pareto curve at inference time. No re-training is required to adjust for different use-case requirements.
230 - 📝 **Punctuation & Capitalization:** Built-in support for punctuation and capitalization in output text.
231
232 ---
233
234 ## Supported Languages
235
236 The model supports **40 language-locales** in total, across three tiers:
237
238 - **Transcription-ready (19 locales):** highest-accuracy ASR, ready out of the box.
239 - **Broad-coverage (13 locales):** production ASR across an additional 13 locales.
240 - **Adaptation-ready (8 locales):** recognized by the tokenizer; fine-tune on in-domain data to unlock full transcription.
241
242 | Tier | Languages (locales) |
243 | :--- | :--- |
244 | **Transcription-ready (19 locales)** | English (en-US, en-GB), Spanish (es-US, es-ES), French (fr-FR, fr-CA), Italian (it-IT), Portuguese (pt-BR, pt-PT), Dutch (nl-NL), German (de-DE), Turkish (tr-TR), Russian (ru-RU), Arabic (ar-AR), Hindi (hi-IN), Japanese (ja-JP), Korean (ko-KR), Vietnamese (vi-VN), Ukrainian (uk-UA) |
245 | **Broad-coverage (13 locales)** | Polish (pl-PL), Swedish (sv-SE), Czech (cs-CZ), Norwegian Bokmål (nb-NO), Danish (da-DK), Bulgarian (bg-BG), Finnish (fi-FI), Croatian (hr-HR), Slovak (sk-SK), Mandarin (zh-CN), Hungarian (hu-HU), Romanian (ro-RO), Estonian (et-EE) |
246 | **Adaptation-ready (8 locales)** | Greek (el-GR), Lithuanian (lt-LT), Latvian (lv-LV), Maltese (mt-MT), Slovenian (sl-SI), Hebrew (he-IL), Thai (th-TH), Norwegian Nynorsk (nn-NO) |
247
248 > **Note:** Transcription-ready and broad-coverage locales (**32 total**) produce ASR transcription out of the box; adaptation-ready locales require fine-tuning on in-domain data to enable full transcription. The model supports uppercase and lowercase letters, punctuation, spaces, and apostrophes.
249
250 > **Note:** We would recommend [Nemotron ASR Streaming (English)](https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b) model for English-only transcription use cases. For all other transcription ready locales, we recommend Nemotron 3.5 ASR to leverage its expanded multilingual capabilities.
251
252 > [!Tip]
253 > **Automatic language detection / language tagging:** When run with `target_lang=auto`, the model detects the spoken language and emits the corresponding **language code/tag** in the output following the terminal punctuation. This lets a single deployment transcribe mixed-language traffic and automatically label each utterance with its detected language — no separate language-ID component required.
254
255 ---
256
257 ## Model Architecture
258
259 **Architecture Type:** FastConformer-CacheAware-RNNT with Prompt
260
261 This model consists of a cache-aware streaming Parakeet (FastConformer) encoder with an RNN-T decoder and language-ID prompt conditioning. It is based on the Cache-Aware [\[1\]](#ref-1) FastConformer [\[2\]](#ref-2) architecture with 24 encoder layers and an RNNT (Recurrent Neural Network Transducer) decoder. The cache-aware streaming design enables efficient processing of audio in chunks while maintaining context from previous frames. Unlike buffered inference, this model maintains caches for all encoder self-attention and convolution layers. This enables reuse of hidden states at every streaming step, where cached activations eliminate redundant computations. As a result, there are no overlapping computations; each processed frame is strictly non-overlapping. This model leverages prompts to guide the transcription process, enabling language-specific transcription from a single ASR model through language ID conditioning.
262
263 <p align="center">
264 <img src="model_architecture.png" alt="Nemotron 3.5 ASR architecture: FastConformer encoder and language-ID encoding are concatenated, projected, and fed to the RNNT decoder" width="900"/>
265 </p>
266
267 The language-ID prompt is fused with the acoustic representation as follows:
268
269 - **FastConformer encoder** processes audio into an acoustic embedding of shape (D=512, T).
270 - **Language Encoding** expands a 128-dim one-hot language vector across the time axis → (K=128, T), broadcasting the language identity to every frame.
271 - **Concatenation** along the feature axis → fused tensor (D + K, T).
272 - **Projection layer** maps the fused features to the RNNT decoder.
273
274 **Network Architecture:**
275 - Encoder: Cache-Aware FastConformer with 24 layers
276 - Decoder: RNNT (Recurrent Neural Network Transducer)
277 - Parameters: 600M
278
279 **This model was developed based on [nvidia/nemotron-speech-streaming-en-0.6b](https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b).**
280
281 ---
282
283 ## Results at a Glance
284
285 ASR performance is measured using Word Error Rate (WER) on the **FLEURS** test sets. Accuracy stays strong across both modes and improves as the chunk size grows, while remaining competitive even at the lowest-latency 80ms setting. Full tables are in [Performance](#performance).
286
287 <p align="center">
288 <img src="fleurs_wer_vs_chunk_size.png" alt="FLEURS average WER vs streaming chunk size (LangID vs Auto-detect)" width="900"/>
289 </p>
290
291 <p align="center">
292 <img src="fleurs_langid_vs_auto.png" alt="FLEURS WER by language: LangID vs Auto-detect at 320ms chunk" width="900"/>
293 </p>
294
295 > **Note:** Japanese and Korean are measured using Character Error Rate (CER) rather than WER, as is standard for these languages.
296
297 ---
298
299 ## Throughput & Efficiency
300
301 Despite being **roughly half the size** (0.6B vs. 1.1B), Nemotron 3.5 ASR serves **far more concurrent streams at far lower latency** than the [Parakeet RNNT 1.1B multilingual model](https://build.nvidia.com/nvidia/parakeet-1_1b-rnnt-multilingual-asr), which runs on buffered streaming. The cache-aware streaming design avoids the redundant recomputation of buffered inference, so a single H100 can sustain dramatically higher concurrency at every chunk size — directly lowering the cost per stream in production. At the lowest-latency 80ms setting, Nemotron sustains **~17× more concurrent streams** (240 vs. 14); at the 1120ms setting it sustains **6× more** (2,400 vs. 400). The latency-vs-concurrency curves tell the same story: Nemotron (solid green) holds low final-token latency well past 1,000 parallel requests, while Parakeet RNNT 1.1B (dashed blue) saturates after only a few hundred.
302
303 <p align="center">
304 <img src="throughput_vs_chunk.png" alt="Concurrent streams supported on a single H100: Nemotron ASR streaming vs Parakeet RNNT, across chunk sizes" width="900"/>
305 </p>
306
307 <p align="center">
308 <img src="latency_vs_parallel.png" alt="Median final-token latency vs number of parallel requests on a single H100, Nemotron vs Parakeet RNNT across chunk sizes" width="900"/>
309 </p>
310
311 > Measured on a single NVIDIA H100. Throughput is the number of real-time streams sustainable in parallel; latency is the median final-token latency at a given level of concurrency.
312
313 ---
314
315 ## Explore more from NVIDIA
316
317 For documentation, deployment guides, enterprise-ready APIs, and the latest open models—including Nemotron and other cutting-edge speech, translation, and generative AI—visit the NVIDIA Developer Portal at [developer.nvidia.com](https://developer.nvidia.com/).
318 Join the community to access tools, support, and resources to accelerate your development with NVIDIA's NeMo, Speech NIM, and foundation models.
319
320 - What is [Nemotron](https://www.nvidia.com/en-us/ai-data-science/foundation-models/nemotron/)?
321 - NVIDIA Developer [Nemotron](https://developer.nvidia.com/nemotron)
322 - [NVIDIA Speech NIM](https://docs.nvidia.com/nim/speech/latest/about/index.html)
323 - [NeMo Documentation](https://docs.nvidia.com/nemo-framework/user-guide/latest/nemotoolkit/asr/models.html)
324
325 Also, check out the following NVIDIA speech models:
326 - Nemotron ASR Streaming (English) (Nemotron 3 ASR) - https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b
327 - Multitalker Parakeet Streaming - https://huggingface.co/nvidia/multitalker-parakeet-streaming-0.6b-v1
328 - Parakeet Realtime EOU - https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
329
330 ---
331
332 ## NVIDIA NeMo
333
334 To train, fine-tune or perform inference with this model, you will need to install [NVIDIA NeMo](https://github.com/NVIDIA/NeMo) [\[4\]](#ref-4). We recommend you install it after you've installed Cython and latest PyTorch version.
335
336 ```bash
337 apt-get update && apt-get install -y libsndfile1 ffmpeg
338 pip install Cython packaging
339 pip install git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[asr]
340 ```
341
342 ## How to Use this Model
343
344 The model is available for use in the NeMo Framework, and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.
345
346 ### Loading the Model
347
348 ```python
349 import nemo.collections.asr as nemo_asr
350 asr_model = nemo_asr.models.ASRModel.from_pretrained(model_name="nvidia/nemotron-3.5-asr-streaming-0.6b")
351 ```
352
353 ### Streaming Inference
354
355 You can use the cache-aware streaming inference script from NeMo - [NeMo/examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py](https://github.com/NVIDIA-NeMo/NeMo/blob/main/examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py)
356
357 This is a prompt-conditioned multilingual model: pass the target language with `target_lang` (e.g. `en-US`, `es-ES`, `de-DE`), or use `target_lang=auto` for automatic language detection.
358
359 ```bash
360 cd NeMo
361 python examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py \
362 model_path=<model_path> \
363 dataset_manifest=<dataset_manifest> \
364 batch_size=<batch_size> \
365 target_lang=<lang_id> \ #language key (e.g. en-US) or "auto" for automatic language detection
366 att_context_size="[56,13]" \ #set the second value to the desired right context from {0,1,3,6,13}
367 strip_lang_tags=true \ #true: remove the detected language tag from the text; false: keep it in the output
368 output_path=<output_folder>
369 ```
370
371 **`strip_lang_tags`** controls how the detected language tag is handled in the output. The model appends a language tag (e.g. `<en-US>`) after the transcript's terminal punctuation:
372 - `strip_lang_tags=false` (keep): the tag is left in the output, so you can read the detected language directly from each utterance — useful for mixed-language traffic and language labeling.
373 - `strip_lang_tags=true` (remove): the tag is stripped, leaving only the clean transcript text — useful when you only need the spoken words.
374
375 ### Setting up Streaming Configuration
376
377 Latency is defined by the `att_context_size` param, where att_context_size = `{num_frames_left_context, num_frame_right_context}`, all measured in **80ms frames**:
378
379 * [56, 0]: Chunk size = 1 (1 × 80ms = 0.08s)
380 * [56, 1]: Chunk size = 2 (2 × 80ms = 0.16s)
381 * [56, 3]: Chunk size = 4 (4 × 80ms = 0.32s)
382 * [56, 6]: Chunk size = 7 (7 × 80ms = 0.56s)
383 * [56, 13]: Chunk size = 14 (14 × 80ms = 1.12s)
384
385 Here, chunk size = current frame + right context; each chunk is processed in non-overlapping fashion.
386
387 ### Input(s): <br>
388
389 **Input Type(s):** Audio, Lang ID <br>
390
391 **Input Format(s):** wav, string <br>
392
393 **Input Parameters:** One-Dimensional (1D) for audio and One-Dimensional (1D) for Lang ID <br>
394
395 **Other Properties Related to Input:** Maximum Length in seconds specific to GPU Memory, No Pre-Processing Needed, Mono channel is required.
396
397 By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions. <br>
398
399 ### Output
400
401 **Output Type(s):** Text String in Input Language <br>
402
403 **Output Format(s):** String <br>
404
405 **Output Parameters:** One-Dimensional (1D) <br>
406
407 **Other Properties Related to Output:** No Maximum Character Length, transcribe punctuation and capitalization.
408
409 By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions. <br>
410
411 ---
412
413 ## Software Integration
414
415 **Runtime Engine:** NeMo 26.06
416
417 **Supported Hardware Microarchitecture Compatibility:**
418 - NVIDIA Ampere
419 - NVIDIA Blackwell
420 - NVIDIA Hopper
421 - NVIDIA Jetson
422 - NVIDIA Lovelace
423 - NVIDIA Turing
424 - NVIDIA Volta
425
426 **Supported Operating System(s):**
427 * Linux <br>
428 * Linux 4 Tegra <br>
429
430 The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.<vr>
431
432
433
434 ---
435
436
437 ## Model Version(s):
438 nemotron-3.5-asr-streaming-0.6b-v1 <br>
439
440 ## Training and Evaluation Datasets:
441
442 ### Training Datasets
443
444 It was trained on speech data across 40 language-locales. The training data is a dynamic blend of public and proprietary internal datasets normalized to have spoken forms in text with punctuation and capitalization, including:
445
446
447 - NVIDIA Riva multilingual ASR training set (Proprietary)
448 - NVIDIA Granary [\[3\]](#ref-3)
449 - Multilingual LibriSpeech (MLS)
450 - Mozilla Common Voice
451 - FLEURS
452 - VoxPopuli / Europarl-ASR
453
454 ** Data Modality: Audio <br>
455
456 ** Audio Training Data Size: 10,000 to 1 Million Hours <br>
457
458 ** Data Collection Method by dataset <br>
459 * Human <br>
460
461 ** Labeling Method by dataset <br>
462 * Human <br>
463 * Synthetic: Synthetic labels were generated from an ensemble of ASR models ([NVIDIA Canary](https://build.nvidia.com/nvidia/canary-1b-asr), [Parakeet Multilingual 1.1B RNNT](https://build.nvidia.com/nvidia/parakeet-1_1b-rnnt-multilingual-asr), [Parakeet CTC 1.1B](https://build.nvidia.com/nvidia/parakeet-ctc-1_1b-asr), [OpenAI Whisper](https://huggingface.co/openai/whisper-large-v3), and [FunASR](https://github.com/modelscope/FunASR)), with punctuation and capitalization (PnC) generated from [Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B).
464
465
466
467
468 ### Evaluation Datasets
469
470 The model was evaluated on multilingual ASR benchmarks:
471
472 - FLEURS
473 - Mozilla Common Voice (MCV)
474 - Multilingual LibriSpeech (MLS)
475 - NVIDIA internal multilingual evaluation sets
476
477 ** Data Collection Method by dataset <br>
478 * Human <br>
479
480 ** Labeling Method by dataset <br>
481 * Human <br>
482
483 ---
484
485 ## Performance
486
487 ASR performance is measured using the Word Error Rate (WER). The tables below report WER (%) on the **FLEURS** test sets across configurable streaming chunk sizes, in two modes:
488 - **Language Input (LangID):** the target language is provided to the model.
489 - **Auto-detect:** the model automatically detects the spoken language.
490
491 > **Note:** Japanese, Korean, and Mandarin are evaluated using Character Error Rate (CER) rather than WER, as is standard for these languages.
492 > **Note on text normalization:** WER/CER are computed after text normalization that aligns the reference and hypothesis (e.g., casing, punctuation, numerals, and formatting conventions). Normalization is not perfect across all 40 language-locales, and residual mismatches between normalized text can inflate the reported error rates — actual transcription quality may be somewhat better than the numbers suggest.
493
494 ### Transcription-ready (19 locales)
495
496 _Languages are ordered by accuracy (lowest WER first)._
497
498 <table>
499 <thead>
500 <tr><th rowspan="2" align="left">Language</th><th colspan="5" align="center" style="background-color:#76b900;color:#ffffff">Language Input (LangID)</th><th colspan="5" align="center" style="background-color:#6b7280;color:#ffffff;border-left:2px solid #cbd5e1;">Auto-detect</th></tr>
501 <tr><th align="center" style="background-color:#eef6e0">80ms</th><th align="center" style="background-color:#eef6e0">160ms</th><th align="center" style="background-color:#eef6e0">320ms</th><th align="center" style="background-color:#eef6e0">560ms</th><th align="center" style="background-color:#eef6e0">1.12s</th><th align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">80ms</th><th align="center" style="background-color:#f3f4f6;">160ms</th><th align="center" style="background-color:#f3f4f6;">320ms</th><th align="center" style="background-color:#f3f4f6;">560ms</th><th align="center" style="background-color:#f3f4f6;">1.12s</th></tr>
502 </thead>
503 <tbody>
504 <tr><td align="left">Spanish (es-US, es-ES)</td><td align="center" style="background-color:#eef6e0;">4.87</td><td align="center" style="background-color:#eef6e0;">4.64</td><td align="center" style="background-color:#eef6e0;">4.39</td><td align="center" style="background-color:#eef6e0;">4.26</td><td align="center" style="background-color:#eef6e0;">4.11</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">5.04</td><td align="center" style="background-color:#f3f4f6;">4.82</td><td align="center" style="background-color:#f3f4f6;">4.48</td><td align="center" style="background-color:#f3f4f6;">4.34</td><td align="center" style="background-color:#f3f4f6;">4.13</td></tr>
505 <tr><td align="left">Italian (it-IT)</td><td align="center" style="background-color:#eef6e0;">5.23</td><td align="center" style="background-color:#eef6e0;">4.85</td><td align="center" style="background-color:#eef6e0;">4.83</td><td align="center" style="background-color:#eef6e0;">4.41</td><td align="center" style="background-color:#eef6e0;">4.25</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">5.28</td><td align="center" style="background-color:#f3f4f6;">4.89</td><td align="center" style="background-color:#f3f4f6;">4.84</td><td align="center" style="background-color:#f3f4f6;">4.47</td><td align="center" style="background-color:#f3f4f6;">4.32</td></tr>
506 <tr><td align="left">Portuguese (pt-BR, pt-PT)</td><td align="center" style="background-color:#eef6e0;">6.29</td><td align="center" style="background-color:#eef6e0;">6.10</td><td align="center" style="background-color:#eef6e0;">5.81</td><td align="center" style="background-color:#eef6e0;">5.65</td><td align="center" style="background-color:#eef6e0;">5.48</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">6.41</td><td align="center" style="background-color:#f3f4f6;">6.19</td><td align="center" style="background-color:#f3f4f6;">5.82</td><td align="center" style="background-color:#f3f4f6;">5.57</td><td align="center" style="background-color:#f3f4f6;">5.47</td></tr>
507 <tr><td align="left">Hindi (hi-IN)</td><td align="center" style="background-color:#eef6e0;">8.13</td><td align="center" style="background-color:#eef6e0;">7.97</td><td align="center" style="background-color:#eef6e0;">7.41</td><td align="center" style="background-color:#eef6e0;">7.05</td><td align="center" style="background-color:#eef6e0;">6.81</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">11.47</td><td align="center" style="background-color:#f3f4f6;">10.83</td><td align="center" style="background-color:#f3f4f6;">9.88</td><td align="center" style="background-color:#f3f4f6;">9.26</td><td align="center" style="background-color:#f3f4f6;">8.23</td></tr>
508 <tr><td align="left">Korean (ko-KR)</td><td align="center" style="background-color:#eef6e0;">7.59</td><td align="center" style="background-color:#eef6e0;">7.70</td><td align="center" style="background-color:#eef6e0;">7.27</td><td align="center" style="background-color:#eef6e0;">7.18</td><td align="center" style="background-color:#eef6e0;">7.12</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">8.31</td><td align="center" style="background-color:#f3f4f6;">8.18</td><td align="center" style="background-color:#f3f4f6;">7.81</td><td align="center" style="background-color:#f3f4f6;">7.49</td><td align="center" style="background-color:#f3f4f6;">7.30</td></tr>
509 <tr><td align="left">English (en-US, en-GB)</td><td align="center" style="background-color:#eef6e0;">9.43</td><td align="center" style="background-color:#eef6e0;">8.88</td><td align="center" style="background-color:#eef6e0;">8.27</td><td align="center" style="background-color:#eef6e0;">7.99</td><td align="center" style="background-color:#eef6e0;">7.91</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">9.72</td><td align="center" style="background-color:#f3f4f6;">9.34</td><td align="center" style="background-color:#f3f4f6;">8.84</td><td align="center" style="background-color:#f3f4f6;">8.80</td><td align="center" style="background-color:#f3f4f6;">8.84</td></tr>
510 <tr><td align="left">German (de-DE)</td><td align="center" style="background-color:#eef6e0;">9.81</td><td align="center" style="background-color:#eef6e0;">9.21</td><td align="center" style="background-color:#eef6e0;">8.83</td><td align="center" style="background-color:#eef6e0;">8.42</td><td align="center" style="background-color:#eef6e0;">8.31</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">9.90</td><td align="center" style="background-color:#f3f4f6;">9.37</td><td align="center" style="background-color:#f3f4f6;">8.87</td><td align="center" style="background-color:#f3f4f6;">8.58</td><td align="center" style="background-color:#f3f4f6;">8.22</td></tr>
511 <tr><td align="left">French (fr-FR, fr-CA)</td><td align="center" style="background-color:#eef6e0;">10.97</td><td align="center" style="background-color:#eef6e0;">10.60</td><td align="center" style="background-color:#eef6e0;">9.79</td><td align="center" style="background-color:#eef6e0;">9.45</td><td align="center" style="background-color:#eef6e0;">9.03</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">11.03</td><td align="center" style="background-color:#f3f4f6;">10.60</td><td align="center" style="background-color:#f3f4f6;">9.84</td><td align="center" style="background-color:#f3f4f6;">9.46</td><td align="center" style="background-color:#f3f4f6;">9.02</td></tr>
512 <tr><td align="left">Russian (ru-RU)</td><td align="center" style="background-color:#eef6e0;">10.84</td><td align="center" style="background-color:#eef6e0;">10.73</td><td align="center" style="background-color:#eef6e0;">9.87</td><td align="center" style="background-color:#eef6e0;">9.60</td><td align="center" style="background-color:#eef6e0;">9.17</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">12.47</td><td align="center" style="background-color:#f3f4f6;">12.09</td><td align="center" style="background-color:#f3f4f6;">11.01</td><td align="center" style="background-color:#f3f4f6;">10.57</td><td align="center" style="background-color:#f3f4f6;">10.03</td></tr>
513 <tr><td align="left">Turkish (tr-TR)</td><td align="center" style="background-color:#eef6e0;">12.34</td><td align="center" style="background-color:#eef6e0;">12.33</td><td align="center" style="background-color:#eef6e0;">12.05</td><td align="center" style="background-color:#eef6e0;">11.34</td><td align="center" style="background-color:#eef6e0;">11.17</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">12.61</td><td align="center" style="background-color:#f3f4f6;">12.28</td><td align="center" style="background-color:#f3f4f6;">11.93</td><td align="center" style="background-color:#f3f4f6;">11.51</td><td align="center" style="background-color:#f3f4f6;">11.32</td></tr>
514 <tr><td align="left">Vietnamese (vi-VN)</td><td align="center" style="background-color:#eef6e0;">13.41</td><td align="center" style="background-color:#eef6e0;">12.87</td><td align="center" style="background-color:#eef6e0;">12.29</td><td align="center" style="background-color:#eef6e0;">11.78</td><td align="center" style="background-color:#eef6e0;">11.18</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">13.59</td><td align="center" style="background-color:#f3f4f6;">13.02</td><td align="center" style="background-color:#f3f4f6;">12.40</td><td align="center" style="background-color:#f3f4f6;">12.02</td><td align="center" style="background-color:#f3f4f6;">11.22</td></tr>
515 <tr><td align="left">Dutch (nl-NL)</td><td align="center" style="background-color:#eef6e0;">14.03</td><td align="center" style="background-color:#eef6e0;">13.43</td><td align="center" style="background-color:#eef6e0;">12.17</td><td align="center" style="background-color:#eef6e0;">11.97</td><td align="center" style="background-color:#eef6e0;">11.46</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">14.09</td><td align="center" style="background-color:#f3f4f6;">13.80</td><td align="center" style="background-color:#f3f4f6;">12.62</td><td align="center" style="background-color:#f3f4f6;">12.24</td><td align="center" style="background-color:#f3f4f6;">11.70</td></tr>
516 <tr><td align="left">Japanese (ja-JP)</td><td align="center" style="background-color:#eef6e0;">13.87</td><td align="center" style="background-color:#eef6e0;">12.90</td><td align="center" style="background-color:#eef6e0;">12.22</td><td align="center" style="background-color:#eef6e0;">11.91</td><td align="center" style="background-color:#eef6e0;">11.48</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">14.97</td><td align="center" style="background-color:#f3f4f6;">13.85</td><td align="center" style="background-color:#f3f4f6;">13.00</td><td align="center" style="background-color:#f3f4f6;">12.38</td><td align="center" style="background-color:#f3f4f6;">11.66</td></tr>
517 <tr><td align="left">Arabic (ar-AR)</td><td align="center" style="background-color:#eef6e0;">13.17</td><td align="center" style="background-color:#eef6e0;">12.65</td><td align="center" style="background-color:#eef6e0;">12.55</td><td align="center" style="background-color:#eef6e0;">12.13</td><td align="center" style="background-color:#eef6e0;">12.03</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">13.47</td><td align="center" style="background-color:#f3f4f6;">12.85</td><td align="center" style="background-color:#f3f4f6;">12.67</td><td align="center" style="background-color:#f3f4f6;">12.18</td><td align="center" style="background-color:#f3f4f6;">12.06</td></tr>
518 <tr><td align="left">Ukrainian (uk-UA)</td><td align="center" style="background-color:#eef6e0;">15.70</td><td align="center" style="background-color:#eef6e0;">15.21</td><td align="center" style="background-color:#eef6e0;">14.55</td><td align="center" style="background-color:#eef6e0;">13.67</td><td align="center" style="background-color:#eef6e0;">13.07</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">18.81</td><td align="center" style="background-color:#f3f4f6;">17.96</td><td align="center" style="background-color:#f3f4f6;">16.79</td><td align="center" style="background-color:#f3f4f6;">15.60</td><td align="center" style="background-color:#f3f4f6;">14.59</td></tr>
519 <tr><td align="left"><strong>Average</strong></td><td align="center" style="background-color:#eef6e0;"><strong>10.38</strong></td><td align="center" style="background-color:#eef6e0;"><strong>10.00</strong></td><td align="center" style="background-color:#eef6e0;"><strong>9.49</strong></td><td align="center" style="background-color:#eef6e0;"><strong>9.12</strong></td><td align="center" style="background-color:#eef6e0;"><strong>8.84</strong></td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;"><strong>11.14</strong></td><td align="center" style="background-color:#f3f4f6;"><strong>10.67</strong></td><td align="center" style="background-color:#f3f4f6;"><strong>10.05</strong></td><td align="center" style="background-color:#f3f4f6;"><strong>9.63</strong></td><td align="center" style="background-color:#f3f4f6;"><strong>9.21</strong></td></tr>
520 </tbody>
521 </table>
522
523 ### Broad-coverage (13 locales)
524
525 _Languages are ordered by accuracy (lowest WER first)._
526
527 <table>
528 <thead>
529 <tr><th rowspan="2" align="left">Language</th><th colspan="5" align="center" style="background-color:#76b900;color:#ffffff">Language Input (LangID)</th><th colspan="5" align="center" style="background-color:#6b7280;color:#ffffff;border-left:2px solid #cbd5e1;">Auto-detect</th></tr>
530 <tr><th align="center" style="background-color:#eef6e0">80ms</th><th align="center" style="background-color:#eef6e0">160ms</th><th align="center" style="background-color:#eef6e0">320ms</th><th align="center" style="background-color:#eef6e0">560ms</th><th align="center" style="background-color:#eef6e0">1.12s</th><th align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">80ms</th><th align="center" style="background-color:#f3f4f6;">160ms</th><th align="center" style="background-color:#f3f4f6;">320ms</th><th align="center" style="background-color:#f3f4f6;">560ms</th><th align="center" style="background-color:#f3f4f6;">1.12s</th></tr>
531 </thead>
532 <tbody>
533 <tr><td align="left">Polish (pl-PL)</td><td align="center" style="background-color:#eef6e0;">19.88</td><td align="center" style="background-color:#eef6e0;">18.92</td><td align="center" style="background-color:#eef6e0;">17.48</td><td align="center" style="background-color:#eef6e0;">16.61</td><td align="center" style="background-color:#eef6e0;">15.15</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">22.65</td><td align="center" style="background-color:#f3f4f6;">21.63</td><td align="center" style="background-color:#f3f4f6;">20.05</td><td align="center" style="background-color:#f3f4f6;">18.52</td><td align="center" style="background-color:#f3f4f6;">16.55</td></tr>
534 <tr><td align="left">Norwegian Bokmål (nb-NO)</td><td align="center" style="background-color:#eef6e0;">20.43</td><td align="center" style="background-color:#eef6e0;">20.07</td><td align="center" style="background-color:#eef6e0;">18.90</td><td align="center" style="background-color:#eef6e0;">18.44</td><td align="center" style="background-color:#eef6e0;">18.10</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">20.91</td><td align="center" style="background-color:#f3f4f6;">20.19</td><td align="center" style="background-color:#f3f4f6;">19.29</td><td align="center" style="background-color:#f3f4f6;">18.76</td><td align="center" style="background-color:#f3f4f6;">18.01</td></tr>
535 <tr><td align="left">Finnish (fi-FI)</td><td align="center" style="background-color:#eef6e0;">21.19</td><td align="center" style="background-color:#eef6e0;">20.57</td><td align="center" style="background-color:#eef6e0;">20.05</td><td align="center" style="background-color:#eef6e0;">18.94</td><td align="center" style="background-color:#eef6e0;">18.34</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">21.61</td><td align="center" style="background-color:#f3f4f6;">20.88</td><td align="center" style="background-color:#f3f4f6;">20.40</td><td align="center" style="background-color:#f3f4f6;">19.36</td><td align="center" style="background-color:#f3f4f6;">18.72</td></tr>
536 <tr><td align="left">Mandarin (zh-CN)</td><td align="center" style="background-color:#eef6e0;">20.56</td><td align="center" style="background-color:#eef6e0;">20.22</td><td align="center" style="background-color:#eef6e0;">20.03</td><td align="center" style="background-color:#eef6e0;">19.51</td><td align="center" style="background-color:#eef6e0;">19.28</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">22.45</td><td align="center" style="background-color:#f3f4f6;">21.07</td><td align="center" style="background-color:#f3f4f6;">20.59</td><td align="center" style="background-color:#f3f4f6;">20.40</td><td align="center" style="background-color:#f3f4f6;">19.87</td></tr>
537 <tr><td align="left">Czech (cs-CZ)</td><td align="center" style="background-color:#eef6e0;">24.18</td><td align="center" style="background-color:#eef6e0;">23.20</td><td align="center" style="background-color:#eef6e0;">22.41</td><td align="center" style="background-color:#eef6e0;">21.04</td><td align="center" style="background-color:#eef6e0;">20.41</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">25.81</td><td align="center" style="background-color:#f3f4f6;">25.12</td><td align="center" style="background-color:#f3f4f6;">23.68</td><td align="center" style="background-color:#f3f4f6;">22.55</td><td align="center" style="background-color:#f3f4f6;">21.45</td></tr>
538 <tr><td align="left">Bulgarian (bg-BG)</td><td align="center" style="background-color:#eef6e0;">24.50</td><td align="center" style="background-color:#eef6e0;">23.58</td><td align="center" style="background-color:#eef6e0;">22.80</td><td align="center" style="background-color:#eef6e0;">21.70</td><td align="center" style="background-color:#eef6e0;">20.53</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">28.28</td><td align="center" style="background-color:#f3f4f6;">27.22</td><td align="center" style="background-color:#f3f4f6;">25.54</td><td align="center" style="background-color:#f3f4f6;">24.05</td><td align="center" style="background-color:#f3f4f6;">21.84</td></tr>
539 <tr><td align="left">Slovak (sk-SK)</td><td align="center" style="background-color:#eef6e0;">25.08</td><td align="center" style="background-color:#eef6e0;">24.14</td><td align="center" style="background-color:#eef6e0;">23.73</td><td align="center" style="background-color:#eef6e0;">22.51</td><td align="center" style="background-color:#eef6e0;">21.28</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">27.59</td><td align="center" style="background-color:#f3f4f6;">26.06</td><td align="center" style="background-color:#f3f4f6;">25.61</td><td align="center" style="background-color:#f3f4f6;">24.15</td><td align="center" style="background-color:#f3f4f6;">22.68</td></tr>
540 <tr><td align="left">Swedish (sv-SE)</td><td align="center" style="background-color:#eef6e0;">25.61</td><td align="center" style="background-color:#eef6e0;">24.85</td><td align="center" style="background-color:#eef6e0;">23.63</td><td align="center" style="background-color:#eef6e0;">22.72</td><td align="center" style="background-color:#eef6e0;">22.17</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">26.28</td><td align="center" style="background-color:#f3f4f6;">25.56</td><td align="center" style="background-color:#f3f4f6;">24.18</td><td align="center" style="background-color:#f3f4f6;">23.57</td><td align="center" style="background-color:#f3f4f6;">22.53</td></tr>
541 <tr><td align="left">Croatian (hr-HR)</td><td align="center" style="background-color:#eef6e0;">27.92</td><td align="center" style="background-color:#eef6e0;">27.09</td><td align="center" style="background-color:#eef6e0;">25.79</td><td align="center" style="background-color:#eef6e0;">24.92</td><td align="center" style="background-color:#eef6e0;">23.97</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">32.13</td><td align="center" style="background-color:#f3f4f6;">31.20</td><td align="center" style="background-color:#f3f4f6;">29.65</td><td align="center" style="background-color:#f3f4f6;">28.95</td><td align="center" style="background-color:#f3f4f6;">27.46</td></tr>
542 <tr><td align="left">Romanian (ro-RO)</td><td align="center" style="background-color:#eef6e0;">31.52</td><td align="center" style="background-color:#eef6e0;">30.93</td><td align="center" style="background-color:#eef6e0;">29.04</td><td align="center" style="background-color:#eef6e0;">27.77</td><td align="center" style="background-color:#eef6e0;">25.90</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">34.22</td><td align="center" style="background-color:#f3f4f6;">33.26</td><td align="center" style="background-color:#f3f4f6;">30.97</td><td align="center" style="background-color:#f3f4f6;">29.84</td><td align="center" style="background-color:#f3f4f6;">26.88</td></tr>
543 <tr><td align="left">Estonian (et-EE)</td><td align="center" style="background-color:#eef6e0;">29.95</td><td align="center" style="background-color:#eef6e0;">29.66</td><td align="center" style="background-color:#eef6e0;">28.59</td><td align="center" style="background-color:#eef6e0;">27.37</td><td align="center" style="background-color:#eef6e0;">26.35</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">30.58</td><td align="center" style="background-color:#f3f4f6;">30.09</td><td align="center" style="background-color:#f3f4f6;">28.72</td><td align="center" style="background-color:#f3f4f6;">28.03</td><td align="center" style="background-color:#f3f4f6;">27.19</td></tr>
544 <tr><td align="left">Danish (da-DK)</td><td align="center" style="background-color:#eef6e0;">32.62</td><td align="center" style="background-color:#eef6e0;">31.51</td><td align="center" style="background-color:#eef6e0;">30.00</td><td align="center" style="background-color:#eef6e0;">28.92</td><td align="center" style="background-color:#eef6e0;">27.49</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">33.15</td><td align="center" style="background-color:#f3f4f6;">31.77</td><td align="center" style="background-color:#f3f4f6;">30.22</td><td align="center" style="background-color:#f3f4f6;">29.33</td><td align="center" style="background-color:#f3f4f6;">27.81</td></tr>
545 <tr><td align="left">Hungarian (hu-HU)</td><td align="center" style="background-color:#eef6e0;">32.70</td><td align="center" style="background-color:#eef6e0;">32.03</td><td align="center" style="background-color:#eef6e0;">30.92</td><td align="center" style="background-color:#eef6e0;">29.72</td><td align="center" style="background-color:#eef6e0;">28.68</td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;">33.40</td><td align="center" style="background-color:#f3f4f6;">32.39</td><td align="center" style="background-color:#f3f4f6;">31.49</td><td align="center" style="background-color:#f3f4f6;">30.20</td><td align="center" style="background-color:#f3f4f6;">29.18</td></tr>
546 <tr><td align="left"><strong>Average</strong></td><td align="center" style="background-color:#eef6e0;"><strong>25.86</strong></td><td align="center" style="background-color:#eef6e0;"><strong>25.14</strong></td><td align="center" style="background-color:#eef6e0;"><strong>24.11</strong></td><td align="center" style="background-color:#eef6e0;"><strong>23.09</strong></td><td align="center" style="background-color:#eef6e0;"><strong>22.13</strong></td><td align="center" style="background-color:#f3f4f6;border-left:2px solid #cbd5e1;"><strong>27.62</strong></td><td align="center" style="background-color:#f3f4f6;"><strong>26.65</strong></td><td align="center" style="background-color:#f3f4f6;"><strong>25.41</strong></td><td align="center" style="background-color:#f3f4f6;"><strong>24.44</strong></td><td align="center" style="background-color:#f3f4f6;"><strong>23.09</strong></td></tr>
547 </tbody>
548 </table>
549
550 ### Adaptation-ready languages (fine-tune to enable)
551
552 These **8 language-locales** are recognized by the tokenizer but are not tuned for production transcription out of the box: **Greek (el-GR), Hebrew (he-IL), Lithuanian (lt-LT), Slovenian (sl-SI), Latvian (lv-LV), Maltese (mt-MT), Thai (th-TH), and Norwegian Nynorsk (nn-NO)**. Fine-tuning on in-domain data is recommended to bring them to production quality.
553
554 Check our [blog post](https://huggingface.co/blog/nvidia/fine-tuning-nemotron-35-asr) of **how to fine-tune Nemotron 3.5 ASR to improve these languages**, including before/after results.
555
556
557 ---
558
559 ## Ethical Considerations
560
561 NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
562
563 Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).
564
565 ---
566