A subtitle error can begin before the first word is recognized. Imagine a clip that opens with music, pauses, then switches from spoken Mandarin into singing. A transcription system must decide which stretches belong in the text and where each stretch begins. Cut away a quiet syllable and the recognizer never gets to hear it. Keep an instrumental passage and it receives sound with no words to recover.

That earlier decision is central to FireRedASR2S, released by Xiaohongshu's FireRed team in February 2026. The package combines a recognizer with voice activity detection, language identification and punctuation prediction. The initial weights and inference code arrived on February 12; the larger LLM recognizer followed on February 25. This is a release about the work surrounding transcription as well as the words themselves.[2]

What changed around the recognizer

The first FireRedASR release, on January 24, 2025, established two approaches: an attention-based encoder-decoder, or AED, and a speech encoder connected through an adapter to a language model. The former offered a smaller recognition model; the latter brought language-model capacity to the task.[1]

The March 2026 technical report describes largely unchanged architectures trained on approximately 200,000 hours, up from about 70,000, with broader language, dialect and domain coverage. The AED branch gains timestamps through an added alignment component.[3]

The choice remains consequential. FireRedASR2-AED has roughly a billion parameters, while FireRedASR2-LLM has more than eight billion.[3] The release advertises word-level timestamps and confidence scores for the AED model, alongside transcription of Mandarin, Chinese dialects and accents, English, and mixed Chinese-English speech.[2] For someone editing captions, knowing where a word belongs can be as useful as a small improvement in the average recognition score.

The cut comes first

The public reference implementation makes the sequence inspectable. Its process method detects speech intervals, slices the waveform and sends those slices to the recognizer. It then adds language information and punctuation, assembling sentences and optional word timings into the result. With detection disabled, the whole recording becomes a single input interval. The implementation expects 16 kHz audio.[4]

This gives a developer a concrete place to investigate a missing phrase. Was the sound excluded before recognition, or did the recognizer hear it and choose the wrong text? Those failures demand different corrections. Replacing the language model cannot restore audio that an earlier stage discarded.

The language label deserves similar attention. In this reference path, language identification runs on the audio segments after transcription; its result is attached to the output. It does not automatically select a different recognizer for each detected language.[4] A label can help another application route a recording, but that application must supply the routing behavior.

The support table reinforces the distinction: language identification covers more than 100 languages, while the advertised transcription coverage centers on Chinese and English.[2] Recognizing which language someone speaks does not establish that the system can transcribe it. A multilingual upload service would need to account for that gap explicitly.

Read the scores at their actual scale

The report gives FireRedASR2-LLM a 2.89% character error rate averaged across four Mandarin test sets, versus 3.05% for AED. These are equal-weight averages across AISHELL-1, AISHELL-2's iOS test set, and WenetSpeech's Internet and meeting sets. They describe the tested collections, rather than the probability that an arbitrary subtitle will be correct.[3]

Speech detection has a different unit. Its reported 97.57% F1 comes from frame-level voice-versus-non-voice classification on 9,443 manually annotated FLEURS recordings spanning 102 languages. The neural detectors were compared at a fixed probability threshold of 0.5. That operating point matters: changing the threshold changes the balance between missed speech and false alarms. Commercial ASR baselines can also change behind their APIs.[3]

For the imagined music-and-speech clip, these results leave a useful question open: how does the complete system behave at the transitions? A strong average recognizer score and a strong detector score do not, by themselves, measure whether the beginning of a softly sung line survives intact.

A speedup with a specific workload

The acceleration guide supplies a separate, concrete result. On one NVIDIA H20, it reports decoding the ten-hour AISHELL-1 test set in 60 seconds with TensorRT, compared with 760 seconds using its PyTorch baseline: approximately 12.7 times faster. The guide concerns FireRedASR2-AED and provides both offline inference and Triton server deployment paths.[5]

That is evidence about bulk decoding under the documented setup. It does not establish how quickly a live caption appears after a person speaks. A live service must also wait for enough audio, decide when to close a segment, handle concurrent requests and deliver the result. The published timing does not separately measure those delays.[5]

The practical gain from this release is that more of the path can be examined. An evaluator can retain detected intervals, inspect the transcript against its timing and test representative transitions: a quiet opening, music under speech, a dialect change, a long pause. The next convincing result would connect those component decisions to complete recordings. For captioning, the final test is whether the right words remain attached to the sound that produced them.

Sources

1. FireRed Team, original FireRedASR repository: January–February 2025 release history and the AED and encoder-adapter-LLM approaches.
2. FireRed Team, FireRedASR2S release notes and model support table, pinned to revision 4e7d9aaf; February–March 2026 release milestones and model capabilities.
3. Kaituo Xu and colleagues, “FireRedASR2S: A State-of-the-Art Industrial-Grade All-in-One Automatic Speech Recognition System,” March 11, 2026, version 1; training changes, evaluation definitions and limitations.
4. FireRed Team, reference pipeline implementation, fireredasr2system.py, revision 4e7d9aaf; segmentation, transcription, language labels and output assembly.
5. FireRed Team, “Accelerating FireRedASR2-AED with TensorRT-LLM,” revision 4e7d9aaf; AISHELL-1 timing on a single H20 and deployment modes.
6. xiquinhosilva, “Xintiandi 21731-Shanghai,” October 14, 2016 photograph, Wikimedia Commons; SOHO Fuxing Plaza, CC BY 2.0.
7. Xiaohongshu, official bilingual copyright-contact page; company address at Fuxing SOHO Plaza, Shanghai.