Model-based turn-end detection
The service detects when a caller has finished a turn rather than relying only on a fixed pause threshold, helping reduce premature cut-offs during spoken input.
Kugel Ear 1
Kugel Ear 1 is a speech-to-text service for voice agents that flags uncertain words and returns candidate spellings with probabilities. It also detects when a caller’s turn ends, helping conversational applications clarify names, addresses, codes, and other error-sensitive inputs.
Speech-to-text with uncertainty signals
Kugel Ear 1 is a speech-to-text service for voice agents. Its main distinction is that it reports words it is unsure about instead of forcing the application to treat every transcript token as certain. For an uncertain word, the response can include alternative spellings and their probabilities, giving the agent enough information to ask the caller for clarification.
The service also determines when a caller’s turn has ended. An application streams the call audio, including pauses, and receives a final transcript when Kugel Ear 1 detects the end of the turn. This is intended for conversational systems handling names, addresses, codes, customer numbers, and phone calls where both transcription errors and talking over the caller can affect the interaction.
Key features
Word-level uncertainty alternatives
When recognition is uncertain, the output includes the original word position, candidate spellings, and probabilities—for example, Viktor, Victor, and Wiktor.
Streaming transcription workflow
The application can stream call audio as it arrives, with pauses included, and receive a final transcript when the model determines that the turn has ended.
Structured output for clarification
The documented output separates a final transcript from a subsequent alternatives message, allowing a voice bot to decide when clarification is needed.
Speech recognition for difficult call inputs
The published examples and tests cover conversational speech, phone-line audio, names, addresses, codes, customer numbers, and other spoken inputs where spelling accuracy matters.
Practical use cases
Confirming names and addresses
When a caller dictates a person’s name or street address, the agent can use alternative spellings to confirm the ambiguous part instead of repeating the entire request.
Checking codes and customer identifiers
For customer numbers, codes, phone numbers, and similar values, the application can identify uncertain words or digits and request a targeted correction before continuing.
Handling phone-line conversations
During a phone conversation, turn-end detection can help the agent wait through meaningful pauses and reduce interruptions while a caller is still reading or speaking.
Building clarification-aware voice agents
Developers building voice bots can use the final transcript and uncertainty message as separate inputs to decide whether to continue automatically or ask a clarification question.
Pros and Cons
Pros
- Returns candidate spellings and probabilities for words it considers uncertain.
- Combines streaming transcription with model-based end-of-turn detection.
- The documented workflow includes pauses in the incoming audio and returns a final transcript when the turn ends.
- Published evaluations include phone-line audio and inputs such as names, addresses, codes, and customer numbers.
- The product page reports a median 72 ms from end of input to final transcript in one support-call measurement.
Cons
- The available product evidence does not establish supported SDKs, complete API documentation, deployment options, or integrations.
- Benchmark results are dataset-specific, and the published evidence does not define accuracy or latency for every language, audio condition, or production workflow.
- Teams need application logic that can interpret alternatives and decide when to ask for clarification.
FAQ
What does Kugel Ear 1 return?
Kugel Ear 1 processes streaming call audio and determines when each speaker turn ends. It then returns a final transcript and identifies words it is unsure about, including the spellings considered and their probabilities.
How does an application send audio to Kugel Ear 1?
The source shows an example WebSocket-style stream in which the application sends a 16 kHz audio configuration and audio chunks, including pauses. The product page does not document a complete integration guide or all supported interfaces.
Who is Kugel Ear 1 for?
It is designed for voice-agent workflows in which an application can use uncertainty information to ask a caller for clarification instead of silently choosing a potentially incorrect spelling.
What has Kugel Ear 1 been benchmarked on?
The published benchmark results cover support calls, simulated phone-line audio, names and addresses, codes and numbers, earnings calls, and end-of-turn detection. Results vary by dataset and should not be treated as a guarantee for every deployment.
Quick Facts
- Category
- Speech-to-text for voice agents
- Primary workflow
- Stream call audio, detect turn end, return transcript and uncertainty alternatives
- Output
- Final transcript plus candidate spellings and probabilities for uncertain words
- Example input
- 16 kHz audio stream with pauses included
- Published benchmark areas
- Support calls, phone lines, names and addresses, codes and numbers, earnings calls
- Product site
- kugelaudio.com