16 models · ranked by HuggingFace downloads
by argmaxinc
WhisperKit CoreML is a collection of Whisper speech recognition models exported to Apple's CoreML format by Argmax, enabling on-device ASR on Apple Silicon (iPhone, iPad, Mac) without network calls. The models run via the WhisperKit framework, which handles chunking, VAD, and decoding on-device. Designed for iOS/macOS applications requiring offline transcription.
by pyannote
Pyannote speaker-diarization-3.1 is a complete speaker diarization pipeline from pyannote.audio that answers 'who spoke when' in an audio recording. It segments audio into speaker-homogeneous regions, clusters them by speaker identity using embedding models, and outputs timestamped speaker labels. Used in meeting transcription, podcast editing, and call center analytics.
by openai
Whisper Large-v3-Turbo is a distilled version of Whisper Large-v3, fine-tuned to achieve most of the large model's transcription accuracy at substantially lower inference cost. It supports over 99 languages and maintains the original model's multilingual ASR quality while requiring fewer decoder layers. MIT licensed and directly compatible with HuggingFace's whisper inference pipeline.
by pyannote
A community-supported speaker diarization pipeline from pyannote.audio that segments multi-speaker audio into per-speaker turns. It combines voice activity detection, speaker embedding, and clustering steps into a single callable pipeline.
by openai
Whisper Large-v3 is OpenAI's full-size ASR model supporting 99+ languages, trained on 680,000 hours of multilingual audio. It delivers state-of-the-art transcription accuracy across languages at the cost of significant inference compute. Apache 2.0 licensed. The Large-v3-Turbo variant (a distilled version) provides similar quality at lower cost for most use cases.
by pyannote
A pretrained voice activity detection pipeline from pyannote.audio, identifying speech segments in audio streams. It is trained on AMI, DIHARD, and VoxConverse corpora and outputs timestamped speech/non-speech labels.
by Qwen
Qwen3-ASR 1.7B is Alibaba's 1.7B parameter automatic speech recognition model supporting multiple languages. It is designed as a production-grade ASR model with strong multilingual performance at a compact size.
by MahmoudAshraf
MMS-300M-1130-forced-aligner is Meta's 300M parameter wav2vec2-based model fine-tuned for forced phoneme-level alignment across 1,130 languages. It takes audio and a text transcript as input and outputs word- or phoneme-level timestamps, enabling subtitle synchronization and linguistic documentation at scale. The CC-BY-NC-4.0 license restricts commercial deployment.
by nvidia
Nemotron-3.5-ASR-Streaming-0.6B is NVIDIA's 600M-parameter cache-aware streaming speech recognition model trained in NeMo. It targets real-time transcription scenarios where audio arrives incrementally, supporting multiple languages with low-latency chunk-by-chunk decoding.
by Qwen
Qwen3-ForcedAligner-0.6B is a forced alignment model from the Qwen3 ASR family, designed to align audio segments to text transcripts at the phoneme or word level. At 0.6B parameters it's compact for deployment in audio processing pipelines. Apache-2.0 licensed.
by microsoft
Phi-4-Multimodal-Instruct is Microsoft's compact multimodal model handling text, audio, images, and video in a single instruction-tuned model. Based on Phi-4-Mini, it covers 23 languages and supports speech recognition, speech translation, and visual QA. MIT-licensed — fully permissive for commercial use.
by jonatasgrosman
wav2vec2-large-xlsr-53-german is a German ASR model fine-tuned from Facebook's XLSR-53 large checkpoint on Mozilla Common Voice 6.0. It was contributed during Hugging Face's XLSR fine-tuning week and is listed on the hf-asr-leaderboard, providing a traceable benchmark reference. The Apache 2.0 license and Azure deployment tag make it accessible for both research and production German transcription pipelines.
by nvidia
Parakeet-TDT-0.6B-v2 is NVIDIA's 600M-parameter English ASR model built on the FastConformer architecture with a Token-and-Duration Transducer (TDT) decoder. It was trained on the Granary and NeMo ASR 3.0 datasets and is listed on the Hugging Face ASR leaderboard. The NeMo framework is required for inference, which differs from the standard Transformers pipeline.
by facebook
Wav2Vec2 XLSR-53 fine-tuned on Common Voice for 53-language phoneme recognition using eSpeak labels, producing phoneme sequences rather than word transcriptions. Useful for linguistic and phonetics applications requiring language-agnostic phoneme extraction. Apache-2.0 licensed.
by facebook
HuBERT-Large fine-tuned on LibriSpeech 960h for English automatic speech recognition. HuBERT uses offline clustering of audio features as pseudo-labels during pretraining, achieving strong ASR quality. Apache-2.0 licensed, it's a foundational ASR model from Meta.
by mlx-community
An MLX-format conversion of NVIDIA's Parakeet TDT-CTC 110M, an English ASR model built on the FastConformer architecture and trained by NVIDIA using the NeMo framework. The MLX conversion enables native Apple Silicon inference. Parakeet TDT-CTC uses a Token-and-Duration Transducer with CTC decoding, which provides fast greedy decoding without beam search overhead.