ksingla025's picture
Fix model card — accurate architecture details, sizes, benchmarks
8143185 verified
|
Raw History Blame Contribute Delete
3.96 kB
metadata
license: other
license_name: whissle-inference-only-1.0
license_link: LICENSE
language:
  - multilingual
  - en
  - hi
  - es
  - fr
  - de
  - it
  - gu
  - mr
pipeline_tag: automatic-speech-recognition
tags:
  - nemo
  - asr
  - onnx
  - cpu
  - emotion
  - age
  - gender
  - intent
  - entity_recognition
datasets:
  - MLCommons/peoples_speech
  - fsicoli/common_voice_17_0
  - ai4bharat/IndicVoices
  - facebook/multilingual_librispeech
  - openslr/librispeech_asr
base_model:
  - nvidia/parakeet-ctc-0.6b
library_name: nemo
extra_gated_heading: Access Whissle STT-meta-1B on Hugging Face
extra_gated_description: >
  This model is licensed for inference only — no training, fine-tuning,
  distillation, or reverse engineering permitted. Accept the license to access.
  Automatic approval.
extra_gated_button_content: Agree and access repository
extra_gated_fields:
  First Name: text
  Last Name: text
  Organization: text
  Country: country
  Date of birth: date_picker
  I want to use this model for:
    type: select
    options:
      - Research
      - Education
      - Commercial product
      - Personal project
      - label: Other
        value: other
  I accept the Whissle Inference-Only License Agreement: checkbox
extra_gated_prompt: >-
  By clicking "Agree", you accept the Whissle Inference-Only License Agreement.
  See the LICENSE file for full terms. Key restrictions: INFERENCE ONLY — no
  training, fine-tuning, distillation, model compression, or reverse engineering
  permitted. Free for inference use under 100M MAU. "Powered by Whissle"
  attribution required for redistribution.

Whissle STT-meta-1B

Multilingual speech recognition model with dual-head tag classifier for real-time speaker metadata extraction. Built on Conformer-CTC architecture with 18 encoder layers. Supports 9 languages with age, gender, emotion, and intent detection per utterance.

Model Details

Architecture Conformer-CTC + dual-head tag classifier
Encoder 512-dim, 18 layers, 4x subsampling
Download size ~488 MB
Format NeMo (.nemo) and ONNX (CPU and GPU compatible)
Sample rate 16 kHz mono
Languages English, Hindi, Spanish, French, German, Italian, Gujarati, Marathi
Base model nvidia/parakeet-ctc-0.6b

Tag Classifier Outputs

Category Classes Labels
Age 6 0-18, 18-30, 30-45, 45-60, 60+, NONE
Emotion 8 NEUTRAL, HAPPY, SAD, ANGRY, FEAR, SURPRISE, DISGUST, NONE
Gender 4 MALE, FEMALE, OTHER, NONE
Intent 10 COMMAND, DESCRIBE, EXCLAIM, EXPLAIN, INFORM, OPINION, QUESTION, REQUEST, STATEMENT, NONE

Quick Start

Use with the Whissle STT Inference Server (ONNX, CPU):

git clone https://github.com/WhissleAI/whissle_stt_inference.git
cd whissle_stt_inference
./setup.sh --model en-meta

Or load directly with NeMo:

import nemo.collections.asr as nemo_asr

asr_model = nemo_asr.models.ASRModel.from_pretrained("WhissleAI/STT-meta-1B")
transcriptions = asr_model.transcribe(["/path/to/your/audio.wav"])

Also usable with PromptingNemo.

Performance

Tested on CPU (Apple M-series):

Audio length Inference time RTF Tags
25.9s 3.6s 0.14x Female, 30-45, Neutral, Describe
1.1s 0.46s 0.42x Female, 18-30, Happy, Question

License

Whissle Inference-Only License — inference only, no training/fine-tuning/distillation/reverse engineering. Free under 100M MAU.

Citation

@misc{whissle2026sttmeta1b,
  title={Whissle STT-meta-1B: Multilingual ASR with Intent, Emotion, and Voice Biometrics},
  author={Whissle AI},
  year={2026},
  url={https://huggingface.co/WhissleAI/STT-meta-1B}
}