Krisp selected Nebius to train foundational voice-conversion models for Accent Localization. Its foundational model took about two weeks to train on one H100 GPU, and the team tested many versions. Nebius reports that a single H100 completed training 50–80% faster than the A100 GPU Krisp had previously used. Nebius provided direct-attached SSDs to move training data between GPUs.
Training datasets usually totalled 2–2.5 terabytes, combining third-party data, internally synthesised data and cleaned public data. Krisp used data augmentation to prepare models for noisy call centres. Nebius’s high-speed SSDs supported data preparation during training to keep GPUs utilised. Krisp evaluated speech quality through subjective testing and objective measures, including word error rate and phoneme error rate.
Krisp then developed compact models designed to run on call-centre computers, with final voice processing on the device. The Nebius customer story describes a three-month collaboration with one planned downtime, communicated in advance. With support from a Nebius cloud solutions architect, Krisp transferred data and started training within a day of the initial discussion. Nebius reports that Krisp met its tight product research and development timeline. Krisp requested an H100 VM option with more CPU power, which Nebius said it would consider. At the time, Krisp planned to keep optimising its real-time models and was considering more accent types. The team also identified more complex distributed training as a challenge for making fuller use of cloud GPU clusters.
Nebius describes Krisp training foundational accent-localization models using H100 GPUs and direct-attached SSDs. Nebius reports that Krisp met its tight product research and development timeline. Krisp requested an H100 VM option with more CPU power, which Nebius said it would consider.
Nebius Customer story24 Nov 2024 · Nebius · Web pageNebius describes Krisp training foundational accent-localization models using H100 GPUs and direct-attached SSDs. Nebius reports that Krisp met its tight product research and development timeline. Krisp requested an H100 VM option with more CPU power, which Nebius said it would consider.
Krisp develops AI audio software for professionals, teams, contact centers, and developers building voice applications. Its AI Meeting Assistant combines background noise and echo removal with meeting recording, transcription, summaries, and action items. Its Call Center AI products provide real-time accent conversion, voice translation, noise cancellation, and agent assistance. Developers can integrate Krisp’s voice models through SDKs and APIs, including tools for isolating speech in voice-agent conversations and translating human calls.
Krisp introduces Make integration for meeting workflows
Krisp announced a beta integration with Make that triggers workflows when meeting notes, transcripts, action items or recordings are ready. It is available on all plans.
Krisp introduced a Fraud Detection beta that identifies tactics such as manufactured urgency and attempts to bypass verification, then alerts agents with a recommended next step. The release also adds a direct Genesys Cloud integration in beta.
USPTO advances Krisp challenges to six Sanas patents
Krisp announced that the USPTO instituted proceedings on all six patents Sanas asserted against it. Four inter partes reviews and two post-grant reviews will proceed before the Patent Trial and Appeal Board, with final validity decisions still pending.
Krisp selected Nebius to train foundational voice-conversion models for Accent Localization. Its foundational model took about two weeks to train on one H100 GPU, and the team tested many versions. Nebius reports that a single H100 completed training 50–80% faster than the A100 GPU Krisp had previously used. Nebius provided direct-attached SSDs to move training data between GPUs.
Training datasets usually totalled 2–2.5 terabytes, combining third-party data, internally synthesised data and cleaned public data. Krisp used data augmentation to prepare models for noisy call centres. Nebius’s high-speed SSDs supported data preparation during training to keep GPUs utilised. Krisp evaluated speech quality through subjective testing and objective measures, including word error rate and phoneme error rate.
Krisp then developed compact models designed to run on call-centre computers, with final voice processing on the device. The Nebius customer story describes a three-month collaboration with one planned downtime, communicated in advance. With support from a Nebius cloud solutions architect, Krisp transferred data and started training within a day of the initial discussion. Nebius reports that Krisp met its tight product research and development timeline. Krisp requested an H100 VM option with more CPU power, which Nebius said it would consider. At the time, Krisp planned to keep optimising its real-time models and was considering more accent types. The team also identified more complex distributed training as a challenge for making fuller use of cloud GPU clusters.
Nebius describes Krisp training foundational accent-localization models using H100 GPUs and direct-attached SSDs. Nebius reports that Krisp met its tight product research and development timeline. Krisp requested an H100 VM option with more CPU power, which Nebius said it would consider.
Nebius Customer story24 Nov 2024 · Nebius · Web pageNebius describes Krisp training foundational accent-localization models using H100 GPUs and direct-attached SSDs. Nebius reports that Krisp met its tight product research and development timeline. Krisp requested an H100 VM option with more CPU power, which Nebius said it would consider.
Krisp develops AI audio software for professionals, teams, contact centers, and developers building voice applications. Its AI Meeting Assistant combines background noise and echo removal with meeting recording, transcription, summaries, and action items. Its Call Center AI products provide real-time accent conversion, voice translation, noise cancellation, and agent assistance. Developers can integrate Krisp’s voice models through SDKs and APIs, including tools for isolating speech in voice-agent conversations and translating human calls.
Krisp introduces Make integration for meeting workflows
Krisp announced a beta integration with Make that triggers workflows when meeting notes, transcripts, action items or recordings are ready. It is available on all plans.
Krisp introduced a Fraud Detection beta that identifies tactics such as manufactured urgency and attempts to bypass verification, then alerts agents with a recommended next step. The release also adds a direct Genesys Cloud integration in beta.
USPTO advances Krisp challenges to six Sanas patents
Krisp announced that the USPTO instituted proceedings on all six patents Sanas asserted against it. Four inter partes reviews and two post-grant reviews will proceed before the Patent Trial and Appeal Board, with final validity decisions still pending.
