Quail Voice Focus 2.3 is a primary-speaker isolation model. It sits in front of speech-to-text (STT) or speech-to-speech (S2S) and keeps only the person speaking to your voice agent, removing background voices and noise in real time. Without it, those voices get transcribed too meaning they end up in the LLM's context, trigger false turns and make the agent answer someone it isn't talking to.
Every call is different
Striking a balance between voice isolation and retention is key. People reach voice agents in every possible setting: on a landline, with a mobile held to the ear or on speakerphone, in a carpeted office or an echoey bathroom, speaking up or barely above a whisper. Sometimes someone else is talking in the background but often nobody is.
A model that removes every voice that is quiet or sounds like background speech – be it quiet, reverberant or far from the mic – also deletes callers who are alone and simply speak quietly or from across the room. Those deletions and lost turns have a heavier than cost than insertions since downstream STT, LLM and S2S systems can account for an extra out-of-place word from a background voice, but a word that was never passed on cannot be recovered.
Voice Focus 2.3 keeps the most dominant voice of those it hears, even when that speaker is quiet, distant or in a reverberant room. The model adapts as the call unfolds, keeping the first speaker but switching whichever speaker is dominant over time.
What's new in version 2.3
Quiet, distant and soft voices are kept. A single speaker talking from across a room, on speakerphone or in a reverberant setting is identified as the primary speaker.
Steady when the voice drops. Callers who suddenly speak more quietly, turn away or walk around keep being transcribed.
Word onsets stay intact. Soft leading syllables are kept, so "spicy" stays "spicy" instead of becoming "icy" and "hair" isn't transcribed as "air".
Holds up in adverse conditions. In very noisy and distorted audio, such as drive-through intercoms and in strongly overlapping speech, the primary speaker is retained more accurately.
Background speakers are removed. Other voices on the call are suppressed, so the transcript holds what your caller said and not what the people around them said.
Benchmarks across nine STT providers
We measured word error rate (WER) in two scenarios: one with interfering speech, and one where a single primary speaker is quiet, distant or in a reverberant room. Each internal test set was streamed in real time through nine STT providers.
Calls with background speakers: 282 real calls on phones and earbuds with other people talking nearby.
One speaker, nothing to remove: 501 clips of a lone speaker, often distant or in a reverberant room.
Word error rate (WER) across 9 live STT providers
DeletedSubstitutedInserted
Each bar reads left to right: deleted, substituted, inserted.
Average of 9 providers · WER in % · lower is better
Raw audio
Voice Focus 2.2 L
Voice Focus 2.3 L
0102030
Raw audio
Voice Focus 2.2 L
Voice Focus 2.3 L
0102030
On calls with background speakers present, insertions from background voices account for 24.9 percentage points of the raw audio's 33.2% WER. While both versions of Voice Focus cut WER roughly in half, version 2.3 gets there with less than half the deletions of version 2.2, accepting more insertions as a trade-off.
On single-speaker audio, version 2.2 more than doubled the WER of the raw audio, almost entirely through deletions. The new Quail Voice Focus 2.3 brings WER down from 26.1% to 12.1%, with deletions close to the raw-audio level. It lowers WER on single-speaker audio in the case of all nine STT providers and on calls for seven of the nine.
Listen for yourself
Ground truth
What is the current interest rate on savings accounts? What is the latest checkout time available?
Raw0:00 / 0:28
Transcript
What is the current interest rate on savings accounts? In my apartment complex yesterday, you would think it was like Super Bowl day. The amount of cheers and everything coming from all the watch parties that was happening, everyone is so excited about the Masters. The Masters Sunday? Yeah. Meanwhile, we're troubled for Congressman Eric Swalwell, as admitted, I think DA takes action against him. What is the latest checkout time available?
Quail Voice Focus 2.3 L0:00 / 0:28
Transcript
What is the current interest rate on savings accounts? What is the latest checkout time available?
- WER, raw
- 350%
- WER, Voice Focus 2.3
- 0%
- Background words removed
- 56
Averages hide what a deletion sounds like. The qualitative examples page has more recordings for both scenarios, each played as the original and with Voice Focus 2.3 L, with spectrograms and AssemblyAI Universal-3.6 Pro transcripts that mark each missed and incorrectly transcribed word.
Calls with an interfering speaker. Calls from our test set with a caller talking to a voice agent while other people speak nearby. Listen for the background voices and note in the original they end up in the transcript as wrong words. Voice Focus 2.3 removes them and keeps the caller's words.
Far-field speech. Calls from our test set with a single quiet speaker at a distance from the microphone. Listen for the quiet speaker: Voice Focus 2.3 keeps them throughout.
Good to know when you upgrade
Far fewer deletions, slightly more insertions. Compared with Voice Focus 2.2, the new model is more permissive: it keeps more speech, so a little more background speech can get through when other voices are loud. This is a deliberate trade-off: deleted words cannot be recovered, while STT, LLM and S2S systems often handle insertions without degrading the user interaction.
Background voices can pass through before the primary speaker starts speaking. Voice Focus decides in real time who the primary speaker is. A voice that speaks before your caller is kept at first - since it could be the primary speaker - then suppressed once the more dominant speaker starts.
Calibrate and raise the enhancement level as needed. 2.3 keeps quiet speech at enhancement levels where 2.2 deleted it. If you chose a low enhancement level to protect quiet callers, set it back to 0.8 and check a sample of your own calls. If background speakers still get through, test 1.0: it removes more background speech but deletes slightly more words when only one person is talking.
Drop-in upgrade. No API change: update to SDK 0.26.0 (Python
aic-sdk3.4.0), download the new model files (older files do not load in 0.26) and usequail-vf-2.3-l-16khzorquail-vf-2.3-s-16khz.Two model sizes. S and L are both lightweight and run on a single CPU core. Choose S for maximum efficiency and L for the strongest performance.
Smaller model files and a faster runtime. The models now ship with 16-bit weights, which halves the file sizes with the same output (L: 20.9 to 11.3 MB; S: 5.3 to 2.8 MB). New SIMD kernels in the SDK runtime process each audio frame 27–38% faster on x86-64 CPUs.
Optimized for downstream AI. Our models are tuned for the accuracy of downstream STT, LLM and S2S systems rather than for how the audio sounds to a human listener.
Language-agnostic. The models work across languages, with no separate setting per language.
Get started
Quail Voice Focus 2.3 is available now in SDK 0.26.0. If you already run 2.2, switch the model ID and test it on a sample of your own calls.
We’re offering 25% off annual subscriptions for the rest of the year. If you’re looking at enterprise volumes, reach out to partnerships to hear about the discounts available.
