Inflection details Pi’s work on overlapping voices, with public release still ahead

Pi’s next step toward group conversation is taking shape inside its speech pipeline, not arriving as a public feature today. Inflection AI says it has integrated NVIDIA models to distinguish speakers and transcribe overlapping speech. But multilingual streaming remains unfinished, and the company has not set a date for releasing the work to users.

In an October 28 engineering post, Inflection described how it is combining speaker labeling, speech recognition and multilingual transcription for future versions of Pi. Its goal is for Pi to follow a group discussion rather than depend on a sequence of individual turns.

Getting the words attached to the right voice

The starting problem was not simply recognizing speech. Inflection’s existing automatic speech recognition system had been fine-tuned on its data, but it did not distinguish individual speakers when producing text. When people talked over one another, their words could end up combined in a single transcript, according to the company.

Inflection first added NVIDIA Streaming Sortformer to identify which speaker was talking at each moment. That process is called speaker diarization. It provides labels for the voices in a conversation, giving the rest of the speech pipeline information about who is speaking rather than treating the audio as one undifferentiated stream.

It then integrated NVIDIA Multitalker-Parakeet into its speech-recognition system. The model uses those speaker labels to guide transcription. Inflection says this lets it capture each person’s words separately, including when voices overlap. The English-language capability therefore depends on two linked jobs: identifying the speakers and using that information to recognize their words.

That dependency makes labeling errors consequential: even small mistakes can affect how Pi interprets and answers a conversation, Inflection says. The company subsequently integrated NVIDIA Nemotron 3 Diarization and reported improved speaker-labeling accuracy in its AMI-SDM benchmark testing, highlighting its handling of overlapping voices.

An offline model still needs a live connection

The multilingual work introduces a different integration problem. Inflection is testing NVIDIA Nemotron 3.5 Transcribe as a single model for speech recognition across languages. Its open weights give the company a way to adapt the model to Pi’s needs, rather than relying only on the model as supplied.

But the checkpoint Inflection tested performs offline transcription, not live streaming. Its early experiments fed the model shorter audio segments. Engineers are still building the buffering and transcript-handling logic needed to connect that approach to Pi’s streaming pipeline. Integrating multilingual transcription with the speaker-aware system remains ongoing work, rather than a completed part of the disclosed stack.

There is also an operating goal behind the multilingual approach. Inflection says a single model could simplify deploying and maintaining speech recognition across languages. It is comparing the candidate with its current pipeline, including whether it can preserve the speed and responsiveness needed for natural conversation.

Promising tests, with a comparison caveat

Inflection describes its initial short-segment transcription results on AMI-SDM English benchmarks as promising. Those evaluations use word error rate, a measure where lower scores are better. The company explicitly cautions that its comparison between Multitalker-Parakeet and Nemotron 3.5 Transcribe is not like-for-like because the models serve different use cases. Those English tests should not be read as a demonstrated multilingual streaming result.

What Inflection is measuring against its current pipeline

• Transcription accuracy: whether the candidate improves recognition of the words being spoken.
• Response latency: how the candidate affects the time needed to respond.
• Serving cost: the expense of running the speech-recognition system.