Audio8 ASR Infinite vs Voxtral and Nemotron: Who Wins on CER/WER?
Audio8 ASR Infinite's CER/WER scores on Aishell and LibriSpeech, compared against Voxtral-Mini-4B-Realtime and Nemotron streaming ASR.
What is Audio8 ASR Infinite and how does it compare to Voxtral and Nemotron?
Audio8 ASR Infinite is an open-weight streaming speech recognition model released by Edge0, built on top of the Voxtral realtime audio architecture with a Qwen2.5-3B-Instruct decoder. On the model card’s own benchmark table, it posts a Chinese CER of 1.750 on Aishell1/test and 2.893 on Aishell4/test, sharply lower than Voxtral-Mini-4B-Realtime-2602’s 16.795 and 16.456 on the same sets. On English LibriSpeech, the ranking flips: Voxtral edges ahead with a WER of 2.210 (test.clean) and 5.552 (test.other), versus Audio8’s 3.042 and 6.808.
TL;DR
- Audio8 ASR Infinite is a bilingual (Chinese/English) streaming ASR model built on the Voxtral realtime audio tower paired with a Qwen2.5-3B-Instruct decoder, released under Apache 2.0.
- On Aishell Chinese benchmarks, it posts CER scores of 1.750 (Aishell1) and 2.893 (Aishell4), far below Voxtral-Mini-4B-Realtime-2602’s 16.795 and 16.456.
- On LibriSpeech English benchmarks, Voxtral actually wins, with WER of 2.210 and 5.552 versus Audio8’s 3.042 and 6.808 on test.clean and test.other.
- Averaged across the four reported test sets, Audio8 ASR Infinite comes out at 3.623 percent error, well under Voxtral’s 10.253 and Nemotron’s 9.524.
- The model uses a rolling KV cache with exact RoPE re-basing to hold a 30-second native context while still supporting unbounded 24/7 transcription without memory growth or drift.
- A selectable audio clock (80/120/160 ms) and configurable delay (240 to 560 ms) let users trade responsiveness for accuracy, with the benchmark numbers reported at an 80 ms clock and 480 ms delay.
- A semantic VAD head, trained separately from the transcription base, aims to distinguish real end-of-turn from thinking pauses or stutters, something acoustic VAD systems often get wrong.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How does the CER/WER benchmark comparison break down?
The model card reports results across four test sets: Aishell1/test and Aishell4/test for Mandarin (measured in Character Error Rate, CER), and LibriSpeech test.clean and test.other for English (measured in Word Error Rate, WER). All three models were evaluated under streaming conditions, with Audio8 and Voxtral tested at a 480 ms transcription delay and Nemotron’s numbers tagged at 560 ms.
The split is stark. Audio8 ASR Infinite dominates the Chinese sets by a wide margin, roughly 9 to 10 times lower CER than Voxtral on Aishell1 and Aishell4. On the English sets, Voxtral is the stronger model, beating Audio8 by less than a full point of WER on both test.clean and test.other. Nemotron’s streaming ASR sits in the middle on Chinese (12.927 and 14.677 CER) but trails on both English sets (3.353 and 7.140 WER), landing behind both Audio8 and Voxtral there.
Averaged across all four sets, Audio8 ASR Infinite reports 3.623 percent error, compared to Voxtral’s 10.253 (calculated over its two comparable sets) and Nemotron’s 9.524. That average is doing a lot of work to mask the fact that Audio8’s advantage is concentrated almost entirely in Mandarin transcription, where its lead is large enough to swing the composite number heavily in its favor.
Why does Audio8 ASR Infinite do so much better on Chinese than English?
The model card doesn’t give an explicit explanation, but the architecture and training details offer a plausible read. Audio8 ASR Infinite inherits its causal audio tower directly from Voxtral Realtime 4B, then trains a new audio projector, frame-length embedding, and swaps in a Qwen2.5-3B-Instruct decoder and LM head. Qwen models are generally strong on Chinese-language tasks by design and training data composition, which likely explains why a Qwen-based decoder outperforms Voxtral’s own decoder stack specifically on Mandarin benchmarks, while giving up some ground on English, where Voxtral’s original decoder may have had an edge to begin with.
This is a useful reminder for anyone benchmarking streaming ASR: architecture family and base decoder choice matter as much as the acoustic front end. A model that inherits a strong audio tower can still under- or over-perform depending on which language the paired decoder was optimized for.
What does “unlimited-length” streaming actually mean here?
Audio8 ASR Infinite’s native context window is 30 seconds, the same limitation most transformer-based streaming ASR models face. What makes it usable for continuous, 24/7 transcription is a rolling KV cache with exact RoPE re-basing. Instead of extending the context indefinitely (which would blow up memory and latency over time), the model slides its cache forward, discarding old audio state while re-basing rotary position embeddings so the model doesn’t lose track of relative timing. The result, according to the model card, is constant memory and constant latency even during continuous operation, rather than a slow drift or a hard cutoff every 30 seconds.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
This matters for real-world deployment. A model that resets or degrades every 30 seconds is unusable for live captioning, call center transcription, or any always-on voice pipeline. A rolling cache design is what turns a short-context checkpoint into something deployable as a long-running service.
What are the audio clock and delay tradeoffs?
Audio8 ASR Infinite exposes a selectable “audio clock”, meaning how often the model makes a token decision: 80 ms, 120 ms, or 160 ms per step, corresponding to 12.5, 8.3, or 6.25 decisions per second. Paired with that is a configurable transcription delay, from 240 ms up to 560 ms depending on the clock chosen. Faster clocks and shorter delays give more responsive, lower-latency transcription; slower clocks and longer delays trade some responsiveness for better accuracy.
The model card documents specific “optimized operation points” that were post-trained and validated: an 80 ms clock with 18 left-pad tokens supporting delays of 240/320/480/560 ms, a 120 ms clock with 12 left-pad tokens supporting 240/480 ms, and a 160 ms clock with 9 left-pad tokens supporting 320/480 ms. The published benchmark numbers use the 80 ms clock at a 480 ms delay, which is a mid-range setting rather than the fastest or most accurate extreme available.
Is Audio8 ASR Infinite worth using over Voxtral or Nemotron?
The honest answer depends on the target language and how the deployment weighs latency against accuracy. For Mandarin-heavy applications, Audio8 ASR Infinite’s CER numbers are decisively better than both alternatives shown on the card, by a wide enough margin that it’s not a close call. For English-only pipelines, Voxtral-Mini-4B-Realtime-2602 currently holds a real, if modest, WER advantage on both LibriSpeech test sets.
Beyond raw accuracy, Audio8 ASR Infinite adds engineering features that matter for production streaming: the semantic VAD heads (trained to distinguish real end-of-turn from pauses and stutters across four different time horizons), the selectable clock/delay tradeoff, and the rolling-cache design for continuous operation. Those aren’t reflected in a CER/WER table but are relevant to anyone evaluating it as an actual deployment component rather than a benchmark entry. The model card also flags that this is a “preview release” focused on the transcription base, with a formal release planned to add frame-level semantic perception beyond plain transcription.
Frequently Asked Questions
What decoder does Audio8 ASR Infinite use?
It uses a Qwen2.5-3B-Instruct decoder and LM head, trained from that initialization, paired with a causal audio tower initialized from Voxtral Realtime 4B.
What license is Audio8 ASR Infinite released under?
It’s released under Apache 2.0, according to the license badge on its model card, and is available as open weights on Hugging Face and GitHub.
How does Audio8 ASR Infinite handle audio longer than its context window?
Its native context is 30 seconds, but a rolling KV cache with exact RoPE re-basing lets it keep transcribing indefinitely while holding memory and latency constant, enabling 24/7 operation via its vLLM deployment path.
Does Audio8 ASR Infinite beat Voxtral on every benchmark?
No. It has a large lead on the Chinese Aishell1 and Aishell4 CER benchmarks but trails Voxtral-Mini-4B-Realtime-2602 on both LibriSpeech English WER benchmarks.
What is semantic VAD and why does it matter?
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Semantic VAD is a set of trained classification heads (8 classes, across 0.5/1.0/2.0/3.0 second horizons) meant to tell apart a real end of speech turn from a thinking pause or stutter, something traditional acoustic voice activity detection often misreads, which can cause premature or delayed transcription cutoffs in streaming systems.
