뉴스 및 전체 검색

kotoba-whisper-v2.0

_Kotoba-Whisper_ is a collection of distilled Whisper models for Japanese ASR, developed through the collaboration bewteen Asahi Ushio and Kotoba Technologies. Following the original work of distil-whisper (Robust Knowledge Distillation via Large-Scale Pseudo Labelling), we employ OpenAI's Whisper large-v3 as the teacher model, and the student model consists the full encoder of the teacher large-v3 model and the decoder with two layers initialized from the first and last layer of the large-v3 model. Kotoba-Whisper is **6.3x faster than large-v3**, while retaining as low error rate as the large-v3. As successor of our first model, kotoba-whisper-v1.0, we release ***kotoba-whisper-v2.0*** trained on the `all` subset of ReazonSpeech (the largest speech-transcription paired dataset in Japanese extracted from Japanese TV audio recordings), which amounts 7,203,957 audio clips (5 sec audio with 18 text tokens in average) after those transcriptions more than 10 WER are removed (see WER Filter for detail). The model was trained for 8 epochs with batch size 256 with sampling rate of 16kHz, and the training and evaluation code to reproduce kotoba-whisper is available at https://github.com/kotoba-tech/kotoba-whisper. Kotoba-whisper-v2.0 achieves better CER and WER than the openai/whisper-large-v3 in the in-domain held-out test set from ReazonSpeech, and achieves competitive CER and WER on the out-of-domain test sets including JSUT basic 5000 and the Japanese subset from CommonVoice 8.0 (see Evaluation for detail).

nb-whisper-small

Introducing the **_Norwegian NB-Whisper Small model_**, proudly developed by the National Library of Norway. NB-Whisper is a cutting-edge series of models designed for automatic speech recognition (ASR) and speech translation. These models are based on the work of OpenAI's Whisper. Each model in the series has been trained for 250,000 steps, utilizing a diverse dataset of 8 million samples. These samples consist of aligned audio clips, each 30 seconds long, culminating in a staggering 66,000 hours of speech. For an in-depth understanding of our training methodology and dataset composition, keep an eye out for our upcoming article. While the main models are suitable for most transcription task, we demonstrate how easy it is to change the output of the main model. The following models are trained 250 additional steps from the main models above, and might be suitable for more targetted use cases: - **Verbatim version**: This lower-cased variant is more literal and suitable for tasks requiring detailed transcription, such as linguistic analysis. You can try the models directly through the HuggingFace Inference API, accessible on the right side of this page. Be aware that initially, the model needs to load and will run on limited CPU capacity, which might be slow. To enhance your experience, we are temporarily hosting some models on TPUs for a few days, significantly boosting their performance. Explore these under the **Spaces** section on the Main Page.

qwen3-asr-1.7b-GGUF

Qwen3-ASR-1.7B is a multilingual speech recognition model supporting 30 languages + 22 Chinese dialects, with support for speech, singing voice, and songs with background music. Audio inputs up to 5 minutes. `qwen3-asr-1.7b-q4_k.gguf` was re-baked with the 24-layer **audio encoder kept at Q8_0** (previously Q4_K like the LLM body; ~0.15 GB larger). The earlier Q4_K export could emit an **empty transcript** (or, on longer audio, a repeated phrase) while still reporting success (CrispASR #240) — the same sub-8-bit encoder drift diagnosed in #218. Encoder-output cosine vs the F16 reference (jfk): **0.9632 → 0.9989** min, **0.9913 → 0.9998** mean — i.e. the tower is back to near-lossless and decode no longer degenerates. The Q8_0 and F16 files were never affected. On JFK speech sample (11s, English): - Q4_K: 0.2x realtime on CPU (4 threads) - Perfect transcription accuracy

index-echo-2b-GGUF

Conversion of IndexTeam/Index-Echo-S2TT-2B for the `index-echo` backend in CrispASR. Weights have passed the independent CPU/CUDA checks below. The backend is available on CrispASR main, including C ABI loading, companion downloads and a pinned nightly regression. The released model transcribes Chinese and translates each subtitle into English, Japanese or Spanish. CrispASR preserves the three-line generation recipe and returns bilingual cues with timestamps and punctuation. English JFK transcription/English translation also matches the released Python model in the CPU fixtures; this is a port-parity result, not a language-wide accuracy claim. The 9B checkpoint is not included. Both files of the chosen precision must be in the same directory. The tower's `index_echo.decoder_file` metadata names its decoder; point CrispASR at the tower, not the decoder. The decoder omits the absent MTP branch and contains exactly the 24 text blocks used by the released speech inference path. Vision weights are omitted.

Fun-ASR-Nano-2512

This model is part of the **FunASR** ecosystem — one industrial-grade open-source toolkit for **ASR · VAD · punctuation · speaker diarization · emotion / event · LLM-ASR**. A Star really helps the project (and keeps you updated): > ⚡ **CPU / edge — no GPU, no Python:** run Fun-ASR-Nano as a single self-contained binary via **llama.cpp / GGUF** (like whisper.cpp), with built-in VAD. Prebuilt binaries + one-command model download → Fun-ASR-Nano-GGUF · runtime · guide Fun-ASR is an end-to-end speech recognition large model launched by Tongyi Lab. It is trained on tens of millions of hours of real speech data, possessing powerful contextual understanding capabilities and industry adaptability. It supports low-latency real-time transcription and covers 31 languages. It excels in vertical domains such as education and finance, accurately recognizing professional terminology and industry expressions, effectively addressing challenges like "hallucination" generation and language confusion, achieving "clear hearing, understanding meaning, and accurate writing."

VibeVoice-ASR-HF

**VibeVoice-ASR** is a unified speech-to-text model designed to handle **60-minute long-form audio** in a single pass, generating structured transcriptions containing **Who (Speaker), When (Timestamps), and What (Content)**, with support for **Customized Hotwords** and over **50 languages**. ➡️ **Demo:** VibeVoice-ASR-Demo ➡️ **Report:** VibeVoice-ASR Technical Report - **🕒 60-minute Single-Pass Processing**: Unlike conventional ASR models that slice audio into short chunks (often losing global context), VibeVoice ASR accepts up to **60 minutes** of continuous audio input within 64K token length. This ensures consistent speaker tracking and semantic coherence across the entire hour.

whisper-large-v3-turbo-gguf

GGUF conversions of openai/whisper-large-v3-turbo for use with transcribe.cpp. Ported from upstream commit 41f01f3, pinned 2026-04-25. Validated against the transformers reference at transcribe.cpp commit 0a26478 on 2026-09-13. OpenAI Whisper large-v3-turbo — converted to GGUF for transcribe.cpp. Multilingual transcription and language detection; unlike the full large-v3 model, this turbo variant does not support speech translation. The v3 family adds Cantonese (yue) and uses a 128-bin mel input. Encoder-decoder transformer; 30-second windows with chunked long-form decoding.

granite-speech-4.1-2b

**Model Summary:** Granite Speech 4.1 2B is a compact and efficient speech-language model, specifically designed for multilingual automatic speech recognition (ASR) and bidirectional automatic speech translation (AST) for English, French, German, Spanish, Portuguese and Japanese. The model was trained on 174,000 hours of audio from public corpora for ASR and AST as well as synthetic datasets tailored to support Japanese ASR, keyword-biased ASR and speech translation. Granite Speech 4.1 2B was trained by modality aligning an intermediate checkpoint of granite-4.0-1b-base to speech on publicly available open source corpora containing audio inputs and text targets. Compared to its predecessor granite-4.0-1b-speech, this model has the same parameter count (the new naming convention reflects actual instead of base LLM size) and provides additional capabilities and improvements: * Higher transcription accuracy for multilingual ASR due to a novel dual-head CTC encoder with both graphemic and BPE outputs and frame importance sampling to focus on informative parts of the audio * Punctuation and truecasing for ASR and AST in all languages (including German noun capitalization) with a simple prompt change * Better keyword list biasing capability for enhanced recognition of names, acronyms and technical jargon Two additional model variants explore different capabilities and inference optimization: * granite-speech-4.1-2b-plus adds speaker-attributed ASR and word-level time stamps * granite-speech-4.1-2b-nar introduces a novel non-autoregressive architecture for higher throughput

muscriptor-gguf

GGUF conversions of the MuScriptor multi-instrument music transcription weights by **Mirelo** and **Kyutai**, converted for use with `muscriptor.cpp`, a C++/ggml port of the reference implementation, and with NeuralNote. These are format conversions: safetensors → GGUF, with the bulk weights cast to float16. Nothing was retrained or fine-tuned. All credit for the model belongs to Mirelo and Kyutai — see `NOTICE.md`. `medium` is the default. `small` is the one to reach for on a machine without a usable GPU; `large` is worth it only with one.

midi-transcription-models