CrispASR Info for SE

The issue provides extensive ASR model information but names no target file, test, or entry point in Subtitle Edit. First locate where speech-recognition model information is documented, then confirm which of the supplied details belong there and what final scope is expected before editing.

NVIDIA Parakeet is a family of high-performance automatic speech recognition models developed by NVIDIA in collaboration with Suno.ai, designed primarily around very fast and accurate speech transcription. The family uses NVIDIA's FastConformer architecture and includes CTC, RNNT, and TDT variants, with model sizes ranging from approximately 600 million to 1.1 billion parameters. The original Parakeet models were trained on tens of thousands of hours of speech and were designed to handle different accents, dialects, background noise, silence, and non-speech portions of recordings.
One of the biggest advantages of Parakeet is its extremely high inference speed. The models have demonstrated the ability to transcribe hundreds or even thousands of times faster than real time under NVIDIA's benchmark conditions, making them particularly well suited to large-scale transcription and subtitle-generation workloads. The CTC models emphasize speed and efficient timestamp calculation, while RNNT models generally prioritize transcription accuracy and robustness. The TDT architecture improves efficiency by predicting both speech tokens and their durations, allowing it to skip unnecessary blank frames during decoding.
The Parakeet family has expanded considerably beyond its original English-only models.
Parakeet TDT 0.6B v3 is a multilingual model supporting 25 languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian, and Ukrainian. It can automatically detect the language and transcribe it without requiring the user to specify the language manually.
For English transcription, the newer Parakeet TDT 0.6B v2 is particularly notable for its combination of speed and accuracy. NVIDIA reports a 6.05% word error rate and extremely high throughput, while the model also provides word-level timestamps, punctuation, and song-to-lyrics transcription. This makes it considerably more useful for media and subtitle work than a basic speech recognizer that only outputs un-timestamped text.
NVIDIA also provides Parakeet Unified 0.6B, which combines offline and streaming transcription into a single model. It can operate with latency as low as approximately 160 ms, while supporting both conventional batch transcription and streaming speech recognition. It includes built-in punctuation and capitalization and is intended for applications such as live captioning, voice assistants, and conversational AI.
Another advantage of Parakeet is its relatively modest hardware requirements compared with many large speech-language models. The 600-million-parameter models are considerably smaller than models such as 3B–7B audio-language systems, while still providing extremely high transcription throughput.
For anime and video subtitle transcription, Parakeet has some particularly useful features, especially word-level timestamps, very high processing speed, punctuation, and song/lyrics recognition.
Overall, NVIDIA Parakeet is one of the strongest choices when transcription speed and efficiency are priorities. Its combination of FastConformer architecture, 600M–1.1B parameter models, extremely high throughput, timestamps, streaming support, punctuation, and increasingly broad language coverage makes it excellent for large-scale transcription. For English media, it is particularly compelling, while for Japanese anime, Qwen3-ASR or a dedicated Japanese Parakeet checkpoint would generally be more appropriate.

NVIDIA Canary ASR is a family of multilingual, multitask speech-recognition models developed by NVIDIA as part of its NeMo speech AI platform.
The current flagship model, Canary-1B-v2, contains approximately 1 billion parameters and is built using a FastConformer encoder and Transformer decoder, allowing it to perform both automatic speech recognition (ASR) and automatic speech translation (AST) within the same model.
Canary-1B-v2 supports 25 languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian, and Ukrainian.
For speech recognition, the model can transcribe audio directly into the same language that was spoken, while its translation capabilities allow English ↔ the other 24 supported languages.
One of Canary's major strengths is that it is designed as a multitask model, meaning the same architecture can handle transcription, speech translation, punctuation and capitalization, and timestamp generation rather than requiring completely separate models for each task.
Canary-1B-v2 provides both word-level and segment-level timestamps, which makes it particularly useful for subtitle creation and precise synchronization of transcripts with video.
It also automatically generates punctuation and capitalization, producing cleaner output that requires less post-processing before being used as subtitles or readable transcripts.
NVIDIA also provides smaller and faster Canary variants.
Canary-1B-Flash is a faster 1-billion-parameter model, while Canary-180M-Flash reduces the model to approximately 182 million parameters and can achieve more than 1,200× real-time processing speed under NVIDIA's reported benchmark conditions.
The Flash models support English, German, French, and Spanish, with word- and segment-level timestamps available for those languages.
Canary models can also be used for streaming and chunked inference, making them suitable for both conventional offline transcription and applications that need speech recognized while the audio is being received.
They are primarily designed to run through NVIDIA NeMo, with support for NVIDIA GPU-based inference and deployment through NVIDIA's broader speech AI ecosystem.
For anime and video subtitle work, Canary has some useful characteristics, particularly its word-level timestamps, punctuation, capitalization, fast inference, and multilingual transcription.
The 25-language model is heavily focused on European languages plus Russian and Ukrainian, so models such as Qwen3-ASR, Voxtral, SenseVoice, or ARK-ASR are considerably more appropriate if Japanese transcription is your primary goal.
Overall, Canary ASR is best viewed as NVIDIA's high-performance multilingual ASR/translation family, combining FastConformer-based acoustic processing with a Transformer decoder to provide transcription, translation, timestamps, and text formatting in a single system.
The 1B-v2 model is the most capable general-purpose Canary model, while the 180M Flash version is particularly interesting when extremely high transcription speed and low resource usage are more important than maximum language coverage.

Cohere Transcribe is Cohere’s open-source speech-to-text model designed specifically for automatic speech recognition (ASR), converting spoken audio into written text with an emphasis on high accuracy and efficient performance.
The main model, cohere-transcribe-03-2026, contains approximately 2 billion parameters and supports 14 languages, including English, Japanese, Chinese, Korean, French, German, Spanish, Portuguese, Arabic, and several others. It is built on a Conformer-based encoder-decoder architecture and is designed to achieve a low word error rate while maintaining fast transcription speeds, making it suitable for applications such as transcribing meetings, interviews, podcasts, videos, training material, and other large collections of audio. Audio files can be submitted through Cohere's Audio Transcriptions API in formats such as WAV, MP3, FLAC, OGG, and MPEG, with a maximum file size of 25 MB. One limitation is that the language must currently be specified manually, as the model does not perform automatic language detection, and it does not provide timestamps or speaker diarization. Overall, Cohere Transcribe is aimed at developers and organizations looking for an accurate, multilingual, and production-oriented speech recognition system that can be accessed through an API or deployed in private infrastructure.

FireRedASR is an open-source, industrial-grade speech-to-text system developed primarily for Mandarin Chinese, Chinese dialects, and English transcription. The original FireRedASR family includes two main approaches: FireRedASR-LLM, a large 8.3-billion-parameter model that combines a speech encoder with a large language model to maximize recognition accuracy, and FireRedASR-AED, a smaller 1.1-billion-parameter attention-based encoder-decoder model designed to provide a better balance between performance and computational efficiency.
The system performs particularly well on Mandarin speech recognition benchmarks and is also capable of recognizing Chinese dialects, English speech, and even singing lyrics. Its open-source nature allows it to be run locally rather than relying exclusively on a cloud API, making it useful for developers building transcription services, speech interfaces, video subtitling tools, or other ASR applications. The newer FireRedASR2S system expands this further by combining speech recognition with voice activity detection, language identification, punctuation prediction, timestamp support, code-switching, and support for more than 20 Chinese dialects and accents.
Overall, FireRedASR is a powerful speech-to-text solution, particularly for Chinese-language transcription, with English support and a choice between a large, accuracy-focused LLM model and a more efficient encoder-decoder model.

Fun-ASR-Nano is an open-source, end-to-end speech-to-text model developed by FunAudioLLM for fast and accurate automatic speech recognition. The standard Fun-ASR-Nano model contains approximately 800 million parameters and is trained on tens of millions of hours of real-world speech data.
Its primary checkpoint supports Chinese, English, and Japanese, with particularly strong support for Chinese speech, including seven major dialect groups and 26 regional accents, while also handling a variety of English and Japanese accents.
The model is designed for low-latency transcription and can be used for both offline batch processing and real-time streaming applications. It also includes features such as hotword support, punctuation processing, voice activity detection, timestamps, and speaker diarization when used with the appropriate supporting models.
One of its more unusual strengths is its ability to recognize song lyrics and rap, even in audio containing background music.
A separate Fun-ASR-MLT-Nano version expands language coverage to 31 languages, making the broader Fun-ASR family suitable for multilingual transcription. The model can be run locally using the FunASR framework, PyTorch, vLLM for high-throughput inference, or even GGUF/llama.cpp-based runtimes for more lightweight deployment.
Overall, Fun-ASR-Nano is a powerful open-source alternative to models such as Whisper, especially for users who need strong Chinese dialect recognition, real-time transcription, lyric recognition, or flexible local deployment.

GigaAM is an open-source speech-to-text model family developed by SberDevices, with a focus on high-quality Russian automatic speech recognition and local deployment.
GigaAM is an open-source family of speech and audio models designed for automatic speech recognition, with its strongest performance historically focused on Russian-language transcription.
Developed by Sber’s Salute Developers team, the system uses a Conformer-based architecture and includes several generations of models, ranging from roughly 220 million to 600 million parameters.
The GigaAM family supports different ASR approaches, including CTC and RNN-T, allowing developers to choose between models optimized for different performance and deployment requirements. The newer GigaAM-v3 models were trained on a much larger and more diverse collection of speech, including call-center recordings, music, voice messages, and atypical speech, and add end-to-end transcription with features such as punctuation and text normalization.
GigaAM also supports word-level timestamps, long-form transcription through voice activity detection, model fine-tuning, ONNX export, and deployment with systems such as Triton Inference Server and TensorRT.
In addition to speech recognition, the GigaAM family includes an emotion recognition model.
More recently, the project introduced GigaAM Multilingual, with 220M and 600M model variants trained on approximately 2 million hours of speech across more than 70 languages, expanding its capabilities beyond its original Russian focus.
Overall, GigaAM is a flexible and powerful open-source speech-to-text system, particularly well suited for Russian and several Eurasian languages, while the newer multilingual models provide broader language coverage for local or server-based transcription applications.

GLM-ASR is an open-source speech-to-text model developed for accurate transcription in challenging real-world audio conditions. Its main model, GLM-ASR-Nano-2512, contains approximately 1.5 billion parameters and supports 17 languages, including Mandarin Chinese, English, Cantonese, Japanese, Korean, and several others.
The model is particularly optimized for Chinese speech and dialect recognition, with strong performance on Cantonese and other regional varieties that can be difficult for general-purpose ASR systems.
It is also specifically designed to handle quiet or low-volume speech, allowing it to recognize voices that may be missed or poorly transcribed by other models.
According to its published benchmarks, GLM-ASR-Nano performs competitively against other open-source and commercial speech recognition systems, including strong results on challenging Chinese benchmarks and real-world meeting audio containing noise and overlapping speech.
It supports deployment through frameworks such as Hugging Face Transformers, vLLM, and SGLang, and can also be exposed through OpenAI-compatible transcription APIs for easier integration into existing applications.
Overall, GLM-ASR-Nano is a relatively compact but powerful speech-to-text model, with its biggest strengths being Chinese and dialect transcription, low-volume speech recognition, multilingual support, and strong accuracy in difficult audio environments.

IBM Granite Speech is an open-source family of speech-to-text and speech-language models developed by IBM for efficient, high-quality automatic speech recognition (ASR), with additional support for automatic speech translation (AST).
The Granite Speech family has evolved from the earlier 2B and 8B speech-aware language models into a broader collection of specialized models designed for different deployment needs.
Current variants include the Granite Speech 4.1 2B, which balances transcription and translation capabilities; 4.1 2B Plus, which adds speaker-attributed transcription, timestamps, and keyword prompting for improving recognition of names, acronyms, and technical terms; and the 4.1 2B NAR, a non-autoregressive version designed for lower latency and higher throughput.
IBM has also introduced the much smaller Granite Speech 5.0 Turbo CTC, a 470-million-parameter model optimized for extremely fast transcription and deployment on laptops, smartphones, and other edge devices.
The Granite Speech models support multiple languages, including English, French, German, Spanish, Portuguese, and Japanese, while earlier models also offered speech translation between English and several European and Asian languages.
The models are designed to perform well with different accents, specialized vocabulary, noisy audio, and real-world recording conditions, making them suitable for applications such as meeting transcription, call centers, voice assistants, subtitles, and enterprise speech-processing systems.
Overall, Granite Speech stands out for combining strong transcription accuracy with relatively compact model sizes, specialized variants, and flexible deployment options ranging from powerful servers to edge devices.

Qwen3-ASR is a family of multilingual speech-to-text models developed by Alibaba’s Qwen team, designed for accurate and efficient automatic speech recognition in both offline and real-time streaming applications.
The family currently includes the Qwen3-ASR-0.6B and Qwen3-ASR-1.7B models, giving users a choice between a smaller, more efficient model and a larger model focused on maximum transcription accuracy.
Qwen3-ASR supports automatic language identification and speech recognition across 30 languages, while also recognizing 22 Chinese dialects and regional varieties, giving the system coverage for 52 languages, dialects, and accents in total.
The supported languages are Chinese, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Hungarian, Macedonian, and Romanian.
In addition, Qwen3-ASR supports regional Chinese varieties including Anhui, Dongbei, Fujian, Gansu, Guizhou, Hebei, Henan, Hubei, Hunan, Jiangxi, Ningxia, Shandong, Shaanxi, Shanxi, Sichuan, Tianjin, Yunnan, Zhejiang, Wu, and Minnan, as well as Cantonese accents from Hong Kong and Guangdong.
One of Qwen3-ASR's notable strengths is its ability to handle more than ordinary speech, with support for spoken audio, singing voices, and songs containing background music, making it potentially useful for lyric transcription as well as standard audio transcription.
The models can also perform streaming transcription, allowing text to be generated with low latency as audio is received, while the separate Qwen3-ForcedAligner-0.6B model can generate detailed timestamps by aligning a transcript with its corresponding audio.
Overall, Qwen3-ASR is a highly capable multilingual speech-to-text system that combines automatic language detection, strong Chinese dialect recognition, real-time streaming, timestamp support, and the ability to transcribe both speech and music, making it suitable for applications such as subtitle generation, meetings, interviews, voice assistants, multilingual transcription services, and song lyric recognition.

Mega-ASR is a robust speech-to-text system designed specifically for real-world audio recorded under difficult acoustic conditions, where standard ASR models may struggle with noise, reverberation, distortion, clipping, transmission problems, or distant microphones.
The model is built on the Qwen3-ASR-1.7B foundation model and uses an additional robustness layer trained on approximately 2.6 million audio samples covering 7 major acoustic conditions and 54 compound scenarios, including environmental noise, far-field speech, physical obstruction, echo and reverberation, recording artifacts, electronic distortion, and audio transmission dropout.
Mega-ASR uses an audio-quality router that analyzes incoming audio and automatically decides whether to use the standard Qwen3-ASR recognition path for clean speech or activate its specialized LoRA-based robustness layer for degraded audio.
This approach helps preserve normal transcription accuracy on clean recordings while improving recognition when the audio is heavily degraded.
The model focuses heavily on semantic recovery and keyword reconstruction, helping it recover speech that may be partially obscured or distorted while reducing common ASR problems such as hallucinated text, repeated phrases, missing words, empty outputs, and dropped utterances.
According to the project's published evaluations, Mega-ASR achieved up to roughly 30% relative improvement over strong open-source and commercial ASR systems in particularly difficult acoustic environments.
Because it is based on Qwen3-ASR, Mega-ASR inherits multilingual speech recognition capabilities, with support including Chinese, Cantonese, English, German, Spanish, French, Italian, Portuguese, Russian, Korean, and Japanese.
Overall, Mega-ASR is less focused on being the smallest or fastest general-purpose transcription model and instead specializes in making speech recognition more reliable when the audio is messy, damaged, noisy, distant, or otherwise difficult to understand, making it particularly useful for field recordings, poor-quality videos, surveillance-style audio, calls, and other challenging real-world transcription tasks.

MOSS-Transcribe-Diarize is an end-to-end speech transcription and speaker diarization model designed for long-form, multi-speaker audio.
The main MOSS-Transcribe-Diarize 0.9B model contains approximately 900 million parameters and combines automatic speech recognition, speaker diarization, timestamp generation, and acoustic event awareness into a single model and inference process.
Unlike traditional transcription pipelines that require separate ASR, speaker detection, speaker clustering, and timestamping systems, MOSS can process an audio or video file in a single pass and generate a structured transcript showing what was said, when it was said, and which anonymous speaker said it.
Its output uses speaker labels such as [S01], [S02], and [S03], along with start and end timestamps for each speech segment.
These speaker labels are anonymous identifiers that distinguish speakers within the recording rather than identifying their real-world identities.
MOSS-Transcribe-Diarize supports transcription and diarization across more than 50 languages and can process recordings up to approximately 90 minutes long in a single inference pass, making it particularly suitable for meetings, interviews, podcasts, lectures, phone calls, discussions, and long-form video content.
The model also supports custom prompts and hotwords, allowing users to provide names, technical terminology, acronyms, or other domain-specific vocabulary that may improve transcription accuracy.
In addition to speech transcription and diarization, it can generate acoustic event annotations, giving the output additional information about sounds and events occurring within the recording.
The model also includes a subtitle-oriented web interface for working with transcriptions and long-form media.
Overall, MOSS-Transcribe-Diarize is especially notable because it combines speech-to-text, speaker diarization, timestamps, long-form audio processing, multilingual recognition, hotword prompting, and acoustic event awareness into one unified model, reducing the need to build and connect multiple separate speech-processing systems.
This makes it particularly useful for creating speaker-labeled subtitles, meeting transcripts, interview transcripts, podcast transcriptions, and other applications involving long recordings with multiple people speaking.

Omnilingual ASR project, is an open-source multilingual speech-to-text system developed by Meta AI and designed to provide automatic speech recognition across more than 1,600 languages, including many languages that are poorly supported by conventional ASR systems.
The current model family includes both CTC-based and LLM-based architectures, with model sizes ranging from approximately 300 million to 7 billion parameters.
The CTC models are primarily focused on fast and efficient transcription, with the 300M, 1B, 3B, and 7B variants offering progressively higher resource requirements and generally stronger recognition performance.
The LLM-ASR models use a speech encoder combined with a language-model-based decoder, providing additional language conditioning and improved flexibility, with 300M, 1B, 3B, and 7B variants available.
The newer v2 models improve recognition accuracy over the original versions, while the Unlimited LLM variants are specifically designed to process audio of essentially unlimited length rather than being restricted to shorter audio segments.
One of Omnilingual ASR's biggest advantages is its enormous language coverage.
The project supports 1,600+ languages, with a complete language list provided by Meta, making it particularly useful for languages that have little or no representation in mainstream systems such as Whisper.
The model family also provides a range of deployment options.
For example, the OmniASR-CTC-300M requires roughly 2 GB of VRAM during inference, while the 1B, 3B, and 7B CTC models require approximately 3 GB, 8 GB, and 15 GB respectively.
The LLM-ASR models require considerably more memory, with the 300M through 7B variants requiring approximately 5 GB, 6 GB, 10 GB, and 17 GB of VRAM respectively under the project's benchmark conditions.
The CTC models are also extremely fast, with the project's benchmarks showing the 300M CTC model running at roughly 96× real-time and the 1B version at roughly 48× real-time on an A100 under the stated test conditions.
This makes the smaller CTC variants particularly interesting for applications where high transcription throughput is more important than using the largest possible model.
Overall, Omnilingual ASR is primarily distinguished by its enormous language coverage rather than simply trying to maximize accuracy in a handful of major languages.
Its combination of 1,600+ supported languages, multiple model sizes, fast CTC variants, larger LLM-based models, and unlimited-length transcription models makes it one of the most interesting open-source choices for multilingual transcription, particularly when dealing with languages that are difficult to support with conventional ASR systems.

Kyutai STT is an open-source speech-to-text system developed by the French AI research lab Kyutai, with a primary focus on real-time and low-latency transcription rather than traditional offline batch recognition.
The current family consists of two main models: Kyutai STT 1B EN-FR, which contains roughly 1 billion parameters and supports English and French, and Kyutai STT 2.6B EN, a larger approximately 2.6-billion-parameter model designed specifically for English and optimized for higher transcription accuracy.
A major feature of Kyutai STT is its streaming architecture, which allows the model to transcribe speech while the audio is still being received rather than waiting for the complete recording.
The 1B EN-FR model has approximately 0.5 seconds of delay and includes a built-in semantic voice activity detector (VAD) that can determine when someone is actually speaking, which is particularly useful for voice assistants and interactive applications.
The 2.6B English model has approximately 2.5 seconds of delay and focuses on maximizing recognition accuracy rather than minimizing latency.
Both models can produce word-level timestamps, making Kyutai STT useful for applications such as subtitle generation, live captions, meetings, voice interfaces, and transcription where accurate timing information is important.
The architecture also supports batch processing, allowing the same models to process many audio streams simultaneously rather than being limited to one recording at a time.
Kyutai reports that an H100 GPU can handle approximately 400 simultaneous real-time streams, demonstrating that the system is designed not only for individual local transcription but also for large-scale, low-latency speech services.
Kyutai's broader speech technology also includes Moshi, a 7-billion-parameter full-duplex speech-to-speech foundation model designed for real-time spoken conversations, as well as Hibiki, a simultaneous speech-translation system.
Kyutai STT itself, however, is specifically intended as the speech-recognition component, and it can be combined with an LLM and Kyutai's TTS models to create complete real-time voice assistants such as Unmute.
Overall, Kyutai STT is best characterized as a real-time ASR system rather than a general-purpose multilingual transcription model.
Its biggest strengths are extremely low latency, streaming transcription, word-level timestamps, semantic VAD, high-throughput batching, and the ability to run locally or on production servers, while its major limitation compared with models such as Qwen3-ASR or Omnilingual ASR is its currently much narrower language coverage.

SenseVoice is an open-source multilingual speech-understanding model developed by the FunAudioLLM team, designed to provide not only automatic speech recognition (ASR), but also spoken-language identification, speech emotion recognition, and audio-event detection in a single model.
The primary released model, SenseVoice-Small, contains approximately 234 million parameters and uses a non-autoregressive encoder architecture, allowing it to perform transcription extremely quickly compared with similarly sized Whisper models.
In the project's benchmark setup, SenseVoice-Small can process approximately 10 seconds of audio in 70 ms, making it particularly attractive for low-latency applications and large-scale batch transcription.
The released SenseVoice-Small checkpoint officially supports five languages: Mandarin Chinese, Cantonese, English, Japanese, and Korean.
Although the broader SenseVoice research was trained on more than 400,000 hours of speech and reports coverage of 50+ languages, the publicly released SenseVoice-Small checkpoint is specifically focused on those five languages, so the two figures should not be confused.
One of SenseVoice's biggest advantages is its rich transcription capability.
Along with the recognized text, the model can identify the language being spoken and classify several emotional states, including happy, sad, angry, neutral, fearful, disgusted, and surprised.
It can also detect audio events such as background music, applause, laughter, crying, coughing, and sneezing, allowing the transcription output to contain contextual information that conventional ASR systems normally discard.
SenseVoice also supports inverse text normalization (ITN), which can convert spoken numbers and expressions into more natural written forms, as well as CTC-based timestamps for aligning recognized text with the original audio.
More recently, the FunASR ecosystem has added a diarization workflow that can combine SenseVoice with VAD, speaker models, and punctuation models to produce speaker-attributed transcripts.
For longer recordings, SenseVoice can be combined with FunASR's voice activity detection system to divide the audio into manageable speech segments before transcription, allowing it to process recordings substantially longer than the short segments used for direct model inference.
SenseVoice is also highly flexible for local deployment.
Overall, SenseVoice is particularly interesting when speed and additional audio understanding are important.
Rather than simply producing text, it can provide transcription, language identification, emotion detection, audio-event detection, punctuation/normalization, and timestamps, while remaining considerably smaller and faster than many general-purpose ASR models.
For anime and video subtitle work, its Japanese support, very high inference speed, timestamps, and ability to detect music and other audio events make it especially useful, although its five-language limitation means that models such as Qwen3-ASR or Omnilingual ASR are better choices when very broad multilingual coverage is required.

ARK-ASR is an open-source multilingual speech-to-text model developed by AutoArk, designed to provide strong transcription accuracy while remaining relatively compact and efficient.
The ARK-ASR family currently includes ARK-ASR-0.6B and ARK-ASR-3B, containing approximately 600 million and 3 billion parameters, respectively.
Both models use an autoregressive architecture that combines a Whisper-style audio encoder, an MLP adapter, and a Qwen-based language-model decoder, allowing the system to use both acoustic information and language-model context when generating transcripts.
ARK-ASR supports 19 languages: Chinese, English, German, Japanese, French, Korean, Spanish, Polish, Italian, Romanian, Hungarian, Czech, Dutch, Finnish, Croatian, Slovak, Slovene, Estonian, and Lithuanian.
The 0.6B model is the compact version and was created using teacher-data adaptation and online policy distillation, allowing the smaller model to learn from a stronger teacher while using considerably less training data than some competing large ASR systems.
In the project's published evaluation, ARK-ASR-0.6B achieved an average English WER of 6.55% across seven Open ASR Leaderboard datasets, slightly outperforming the Qwen3-ASR-0.6B baseline at 6.93% while using the same approximate model scale.
The larger ARK-ASR-3B is aimed at maximum recognition performance.
Its published benchmark reports an average 5.04% WER across AMI, Earnings22, GigaSpeech, LibriSpeech, SPGISpeech, and VoxPopuli, along with an RTFx of approximately 491× real time under the reported evaluation conditions.
ARK-ASR also performs particularly well on Chinese and Japanese, making it interesting for multilingual subtitle and video-transcription workflows.
The 3B model reports a Chinese CER of 1.80% on AISHELL-1, while the model family explicitly includes Japanese among its supported languages.
One particularly interesting aspect of ARK-ASR is its accuracy-to-model-size ratio.
The 0.6B model is substantially smaller than many high-performing LLM-based ASR systems, yet its reported benchmark performance is competitive with larger models, while the 3B version pushes accuracy considerably further.
Overall, ARK-ASR is a strong general-purpose multilingual ASR option that combines a Whisper-style acoustic encoder with a Qwen language decoder, supports 19 languages, offers both 0.6B and 3B variants, and provides local as well as high-throughput server deployment options.

Voxtral is Mistral AI’s family of speech and audio-understanding models, designed to provide high-quality speech-to-text transcription while also giving language models the ability to directly understand and reason over audio.
The original Voxtral family includes Voxtral Mini, a roughly 3-billion-parameter model intended for local and edge deployment, and Voxtral Small, a much larger 24-billion-parameter model aimed at production-scale audio understanding. Both can process long-form audio, automatically detect the spoken language, and support tasks beyond simple transcription, including summarization, question answering, speech translation, and function calling directly from spoken instructions.
For dedicated transcription, Mistral now offers Voxtral Mini Transcribe 2, which is optimized specifically for batch speech recognition and supports recordings of up to 3 hours per request. It provides speaker diarization, word-level timestamps, and context biasing, allowing users to supply up to 100 custom terms to improve recognition of names, technical vocabulary, or other specialized words.
The model supports 13 languages: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, and Dutch.
Mistral also provides Voxtral Mini Transcribe Realtime, a roughly 4-billion-parameter open-weight model designed specifically for live transcription. It can transcribe speech as audio arrives with latency configurable to below 200 ms, making it particularly suitable for voice assistants, live captions, interactive applications, and real-time subtitle generation. The realtime model supports the same 13 languages as the dedicated transcription system and can be deployed locally rather than requiring a cloud-only API.
One of Voxtral's major strengths is its ability to retain semantic understanding of the audio rather than treating speech recognition as simply converting audio into text. The larger Voxtral models can answer questions about recordings, summarize conversations, analyze long-form audio, and perform speech translation without requiring a separate transcription model followed by a conventional text LLM.
The original Voxtral models support audio contexts of up to approximately 30 minutes for transcription and 40 minutes for audio understanding, while the newer dedicated transcription model is designed for substantially longer recordings.
For anime and video subtitle work, Voxtral is particularly interesting because the dedicated transcription models provide word-level timestamps and speaker diarization, while Japanese is one of the officially supported languages.
The realtime model is also small enough to make local deployment considerably more practical than very large audio-language models. However, its 13-language coverage is much narrower than models such as Qwen3-ASR or Omnilingual ASR, so Voxtral is more attractive when you prioritize transcription quality, timestamps, diarization, and audio understanding rather than maximum language coverage.
Overall, Voxtral is a versatile speech platform rather than just a conventional ASR model, combining dedicated high-quality transcription models with larger audio-understanding models and a low-latency realtime model. Its strongest features are Japanese support, long-form transcription, word-level timestamps, speaker diarization, context biasing, realtime transcription, and direct audio understanding, making it a particularly capable option for subtitles, meetings, interviews, podcasts, and other long-form media.