File size: 5,883 Bytes
a7c2243 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 | .. _choosing-a-model:
Choosing a Model
================
NeMo offers many pretrained speech models. This guide helps you pick the right one for your use case.
ASR: Which Model Should I Use?
------------------------------
.. list-table::
:widths: 30 25 45
:header-rows: 1
* - I want to...
- Recommended Model
- Why
* - Get the best accuracy on English
- `Canary-Qwen 2.5B <https://huggingface.co/nvidia/canary-qwen-2.5b>`_
- State-of-the-art English ASR. For very fast offline alternatives with almost SOTA accuracy, use `Parakeet-TDT V2 <https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2>`_ or `Parakeet-TDT V3 <https://huggingface.co/nvidia/parakeet-tdt-1.1b>`_.
* - Transcribe multiple languages
- `Canary-1B V2 <https://huggingface.co/nvidia/canary-1b-v2>`_
- Supports 25 EU languages + translation between them. AED decoder.
* - Transcribe European languages (ASR only, auto language detection)
- `Parakeet-TDT 0.6B V3 <https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3>`_
- 25 European languages in one model; automatic language detection; punctuation, capitalization, and word/segment timestamps; long-form and streaming options. No speech-to-text translation—use Canary-1B V2 if you need translation.
* - Stream audio in real-time
- `Nemotron-Speech-Streaming <https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b>`_
- Low-latency streaming English ASR with configurable chunk sizes. Cache-aware FastConformer + RNN-T.
* - Minimize model size
- `Canary-180M Flash <https://huggingface.co/nvidia/canary-180m-flash>`_
- Smallest multilingual model. Good for edge deployment.
* - Use CTC decoding (simpler pipeline)
- `Parakeet-CTC-1.1B <https://huggingface.co/nvidia/parakeet-ctc-1.1b>`_
- Non-autoregressive. Fast. Good with external language models.
* - Integrate with an external LM
- Any Parakeet model + NGPU-LM
- GPU-accelerated n-gram LM fusion for CTC, RNNT, and TDT models.
* - Transcribe multi-speaker meetings
- `Multitalker Parakeet Streaming <https://huggingface.co/nvidia/multitalker-parakeet-streaming-0.6b-v1>`_
- Handles overlapping speech in real-time with speaker-adapted decoding.
TTS: Which Model Should I Use?
------------------------------
.. list-table::
:widths: 30 25 45
:header-rows: 1
* - I want to...
- Recommended Model
- Why
* - Generate high-quality multilingual speech
- `MagpieTTS <https://huggingface.co/nvidia/magpie_tts_multilingual_357m>`_
- End-to-end LLM-based TTS. Supports voice cloning and multiple languages.
* - Fast, controllable English synthesis
- `FastPitch <https://huggingface.co/nvidia/tts_en_fastpitch>`_ + `HiFi-GAN <https://huggingface.co/nvidia/tts_hifigan>`_
- Cascaded pipeline with pitch/duration control. Well-tested.
* - Generate discrete audio tokens
- Audio Codec
- Neural audio codec for tokenizing audio. Used by MagpieTTS internally.
Speaker Tasks: Which Model Should I Use?
-----------------------------------------
.. list-table::
:widths: 30 25 45
:header-rows: 1
* - I want to...
- Recommended Model
- Why
* - Determine who spoke when
- `Streaming Sortformer <https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2.1>`_, `Offline Sortformer <https://huggingface.co/nvidia/diar_sortformer_4spk-v1>`_
- End-to-end diarization for up to 4 speakers. Use streaming for real-time; use offline for batch.
* - Verify/identify a speaker
- `TitaNet <https://huggingface.co/nvidia/speakerverification_en_titanet_large>`_
- Extracts speaker embeddings for verification and identification.
* - Detect voice activity
- `MarbleNet <https://huggingface.co/nvidia/Frame_VAD_Multilingual_MarbleNet_v2.0>`_
- Frame-level VAD. Multilingual. Works as a preprocessing step.
Speech Language Models: Which Model Should I Use?
-------------------------------------------------
.. list-table::
:widths: 30 25 45
:header-rows: 1
* - I want to...
- Recommended Model
- Why
* - Ask questions about audio content
- `Canary-Qwen 2.5B <https://huggingface.co/nvidia/canary-qwen-2.5b>`_ (SALM)
- LLM augmented with speech understanding. Can transcribe, translate, and answer questions about audio.
* - Build a speech-to-speech system
- DuplexS2SModel
- Full-duplex model that both understands and generates speech.
Decision Flowchart
------------------
.. code-block:: text
What do you want to do?
│
├─ Transcribe speech to text (ASR)
│ ├─ Best accuracy on English? → Canary-Qwen 2.5B (or Parakeet-TDT V2/V3 for fast offline)
│ ├─ Multiple languages + translation? → Canary-1B V2
│ ├─ European multilingual ASR (auto LID)? → Parakeet-TDT 0.6B V3
│ ├─ Stream audio in real-time? → Nemotron-Speech-Streaming
│ └─ Multi-speaker meeting? → Multitalker Parakeet Streaming
│
├─ Generate speech from text (TTS)
│ ├─ Multilingual / voice cloning? → MagpieTTS
│ └─ English with pitch control? → FastPitch + HiFi-GAN
│
├─ Identify speakers
│ ├─ Who spoke when? → Streaming Sortformer or Offline Sortformer
│ └─ Verify identity? → TitaNet
│
├─ Enhance audio quality → See Audio Processing models
│
└─ Speech-aware LLM → Canary-Qwen 2.5B (SALM)
Where to Find Models
--------------------
All pretrained NeMo models are available on:
- `HuggingFace Hub (nvidia) <https://huggingface.co/nvidia>`_ — search for "nemo" or specific model names
- `NGC Model Catalog <https://catalog.ngc.nvidia.com/models?query=nemo&orderBy=weightPopularDESC>`_ — NVIDIA's model registry
See :doc:`../checkpoints/intro` for instructions on loading pretrained models.
|