Xiaomi open-sources CocktailASR-1 to pick one voice out of a crowd
Xiaomi just gave away one of the harder pieces of the voice stack. On September 11 the company released and open-sourced Xiaomi-CocktailASR-1, an end-to-end speech recognition model built to solve the cocktail party problem — transcribing one specific person's speech from an audio mix where several people are talking over each other, without a separation stage in front of the model.
The design is unusually direct. You feed the model a short reference clip of the target speaker (one to four seconds) alongside the mixed audio, and it treats the clip as a voiceprint prompt: it transcribes only that person and ignores the rest, rather than first splitting the audio into separate speaker tracks the way classical pipelines do. It is built on an LLM-based architecture, and it keeps two abilities that earlier target-speaker models traded away — single-speaker accuracy comparable to mainstream ASR models (no model switching when the crowd leaves the room) and negative sample rejection, meaning it outputs empty text when the target speaker is absent instead of hallucinating a transcript from the wrong voice. A chain-of-thought mode can also expose its reasoning — speaker count, gender, voiceprint similarity — before committing to the transcript.
The numbers back the claim. On LibriSpeechMix 2mix, the target-speaker word error rate drops from 5.40% for the previous state of the art to 2.90% — a 46% relative reduction — and on LibriMix 2mix it improves from 4.84% to 4.11%. The report is blunt about the competition: top-tier ASR models like Qwen3-ASR-1.7B and StepAudio2 cannot handle overlapping speakers at all, scoring word error rates between 60% and 160% on highly overlapping synthetic sets, while Gemini 2.5 Pro manages 30% to 80% but lacks specialized training for noisy, real-world acoustics.
Why this matters beyond the benchmark table. Smart speakers, wearables, and meeting recorders all live in multi-person environments, and false triggers are the failure users actually notice — the device transcribing a conversation it was never meant to hear. A rejection mechanism baked into the model weights (no extra threshold to tune at inference) is the kind of unglamorous reliability work consumer voice products have needed for a decade. It also follows the open-source pattern that has defined Chinese consumer-AI players this year: ship the weights, let the ecosystem build on them, and set the reference implementation. Worth noting the scale of the training corpus behind it — roughly 600,000 hours of single-speaker TS-ASR data plus dedicated negative-sample sets — which is not something most labs can replicate casually.
What to watch: whether Xiaomi follows up with on-device variants for its phones and wearables, and whether the rejection capability gets adopted as a standard eval for voice agents.
If you build voice products, does baked-in speaker targeting change your stack — or is Whisper-style general transcription still good enough? Tell us in the comments.
