🎉 NEW: Canary 1B V2 is now available!
🌍 25 European Languages | ⏱️ Much Improved Timestamp Prediction | 🔄 Enhanced ASR & AST | 🔗 Try it here: nvidia/canary-1b-v2
Description:
NVIDIA NeMo Canary Flash [1] is a family of multilingual multi-tasking models based on Canary architecture [2] that achieve state-of-the-art performance on multiple speech benchmarks. With 883 million parameters and an inference speed of more than 1000 RTFx (on open-asr-leaderboard datasets), canary-1b-flash supports automatic speech-to-text recognition (ASR) in four languages (English, German, French, Spanish) and translation from English to German/French/Spanish and from German/French/Spanish to English with or without punctuation and capitalization (PnC). Additionally, canary-1b-flash offers an experimental feature for word-level and segment-level timestamps in English, German, French, and Spanish.
This model is released under the permissive CC-BY-4.0 license and is available for commercial use.
Discover more from NVIDIA:
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.
Join the community to access tools, support, and resources to accelerate your development with NVIDIA’s NeMo, Riva, NIM, and foundation models.
Canary is an encoder-decoder model with FastConformer [3] Encoder and Transformer Decoder [4]. With audio features extracted from the encoder, task tokens such as <target language>, <task>, <toggle timestamps> and <toggle PnC> are fed into the Transformer Decoder to trigger the text generation process. Canary uses a concatenated tokenizer [5] from individual SentencePiece [6] tokenizers of each language, which makes it easy to scale up to more languages. The canary-1b-flash model has 32 encoder layers and 4 decoder layers, leading to a total of 883M parameters. For more details about the architecture, please refer to [1].
NVIDIA NeMo
To train, fine-tune or transcribe with canary-1b-flash, you will need to install NVIDIA NeMo.
How to Use this Model
The model is available for use in the NeMo Framework [7], and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.
Input Type(s): Audio Input Format(s): .wav or .flac files Input Parameters(s): 1D Other Properties Related to Input: 16000 Hz Mono-channel Audio, Pre-Processing Not Needed
Input to canary-1b-flash can be either a list of paths to audio files or a jsonl manifest file.
If the input is a list of paths, canary-1b-flash assumes that the audio is English and transcribes it. I.e., canary-1b-flash default behavior is English ASR.
python
1output = canary_model.transcribe(2['path1.wav','path2.wav'],3 batch_size=16,# batch size to run the inference with4 pnc='yes',# generate output with Punctuation and Capitalization5)67predicted_text_1 = output[0].text
8
canary-1b-flash can also generate word and segment level timestamps
For audio files longer than 10 seconds, we recommend using longform inference script (explained in next section) with chunk_len_in_secs=10.0 to generate timestamps.
To use canary-1b-flash for transcribing other supported languages or perform Speech-to-Text translation or provide word-level timestamps, specify the input as jsonl manifest file, where each line in the file is a dictionary containing the following fields:
yaml
1# Example of a line in input_manifest.json2{3"audio_filepath":"/path/to/audio.wav",# path to the audio file4"source_lang":"en",# language of the audio input, set `source_lang`==`target_lang` for ASR, choices=['en','de','es','fr']5"target_lang":"en",# language of the text output, choices=['en','de','es','fr']6"pnc":"yes",# whether to have PnC output, choices=['yes', 'no']7"timestamp":"yes",# whether to output word-level timestamps, choices=['yes', 'no']8}
and then use:
python
1output = canary_model.transcribe(2"<path to input manifest file>",3 batch_size=16,# batch size to run the inference with4)
Longform inference with Canary-1B-flash:
Canary models are designed to handle input audio smaller than 40 seconds. In order to handle longer audios, NeMo includes speech_to_text_aed_chunked_infer.py script that handles chunking, performs inference on the chunked files, and stitches the transcripts.
The script will perform inference on all .wav files in audio_dir. Alternatively you can also pass a path to a manifest file as shown above. The decoded output will be saved at output_json_path.
Note that for longform inference with timestamps, it is recommended to use chunk_len_in_secs of 10 seconds.
Output:
Output Type(s): Text Output Format: Text output as a string (w/ timestamps) depending on the task chosen for decoding Output Parameters: 1-Dimensional text string Other Properties Related to Output: May Need Inverse Text Normalization; Does Not Handle Special Characters
The canary-1b-flash model is trained on a total of 85K hrs of speech data. It consists of 31K hrs of public data, 20K hrs collected by Suno, and 34K hrs of in-house data.
The datasets below include conversations, videos from the web and audiobook recordings.
Canary-1B-Flash is trained using the NVIDIA NeMo Framework [7] for a total of 200K steps with 2D bucketing [1] and optimal batch sizes set using OOMptimizer [8].The model is trained on 128 NVIDIA A100 80GB GPUs.
The model can be trained using this example script and base config.
The tokenizers for these models were built using the text transcripts of the train set with this script.
Inference:
Engine: NVIDIA NeMo Test Hardware :
A6000
A100
V100
Performance
For ASR and AST experiments, predictions were generated using greedy decoding. Note that utterances shorter than 1 second are symmetrically zero-padded upto 1 second during evaluation.
English ASR Performance (w/o PnC)
The ASR performance is measured with word error rate (WER), and we process the groundtruth and predicted text with whisper-normalizer.
WER on HuggingFace OpenASR leaderboard:
Version
Model
RTFx
AMI
GigaSpeech
LS Clean
LS Other
Earnings22
SPGISpech
Tedlium
Voxpopuli
nemo-main
canary-1b-flash
1045.75
13.11
9.85
1.48
2.87
12.79
1.95
3.12
5.63
Inference speed on different systems
We profiled inference speed on the OpenASR benchmark (batch_size=128) using the real-time factor (RTFx) to quantify throughput.
Number of characters per minute on MUSAN 48 hrs eval set
Version
Model
# of character per minute
nemo-main
canary-1b-flash
60.92
Noise Robustness
WER on Librispeech Test Clean at different SNR (signal to noise ratio) levels of additive white noise
Version
Model
SNR 10
SNR 5
SNR 0
SNR -5
nemo-main
canary-1b-flash
2.34
3.69
8.84
29.71
Model Fairness Evaluation
As outlined in the paper "Towards Measuring Fairness in AI: the Casual Conversations Dataset" [9], we assessed the canary-1b-flash model for fairness. The model was evaluated on the CausalConversations-v1 dataset, and the results are reported as follows:
Gender Bias:
Gender
Male
Female
N/A
Other
Num utterances
19325
24532
926
33
% WER
14.66
12.44
17.17
27.56
Age Bias:
Age Group
(18-30)
(31-45)
(46-85)
(1-100)
Num utterances
15956
14585
13349
43890
% WER
13.18
13.45
13.64
13.41
(Error rates for fairness evaluation are determined by normalizing both the reference and predicted text, similar to the methods used in the evaluations found at https://github.com/huggingface/open_asr_leaderboard.)
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.
Please report security vulnerabilities or NVIDIA AI Concerns here.
