LAM (Large Audio Model)

Most voice agents are a chain: speech recognition, then a text language model, then speech synthesis. A large audio model collapses part or all of that chain into one network that takes audio in and produces audio or text out. It matters to buyers because it changes latency, how the agent handles tone and interruption, and what can go wrong.

How it works. The model is trained on paired audio and text at scale, so its internal representation carries both the words and the acoustic signal: pitch, pace, hesitation, background noise. In a speech-to-speech configuration it listens to the caller's audio stream, reasons over it and generates the spoken reply directly, skipping a separate transcription step. In an audio-understanding configuration it emits text (a transcript, an intent, an emotion label) with far more context than a plain recognizer. Tool calls, such as a calendar lookup, still happen through text.

Why it matters for voice agents. Two consequences. First, latency: removing a model boundary removes a serialization step, which is part of how sub-second responses become realistic on a phone line. Second, paralinguistics: a chained system loses tone at the transcription step, so "fine" from a frustrated caller reads the same as "fine" from a happy one. An audio-native model can hear the difference and respond to it. The cost is control: a system that speaks without producing text first is harder to audit, ground against documents and constrain.

Example. A SaaS support team trials two agents on the same billing line. The chained agent transcribes "I've been charged twice" and answers correctly but flatly. The audio-native agent hears the raised voice, acknowledges the frustration in its first sentence, and then runs the same refund lookup. On the transcript both calls look identical; on the recording the second ends calmer. The team keeps the chained agent for the policy-heavy cancellation flow, where every answer must trace to a document, and uses the audio-native one for first contact.

LAM vs LLM. A large language model works on text tokens; everything it knows about a call comes from a transcript. A large audio model works on audio tokens, or on both. The LLM is easier to ground, log and test; the LAM is better at prosody, overlap and emotion. Production stacks increasingly place an audio model for listening and speaking around a text model for reasoning and tool use.

Common pitfalls.

• Assuming audio-native means no transcript is needed; compliance review and CRM notes still need text.
• Harder hallucination detection, because there is no intermediate text to check before the words are spoken.
• Voice and language coverage tied to a single vendor's model, with no way to swap the voice alone.
• Benchmarks run on clean wideband audio instead of compressed phone calls.
• Treating emotion labels as facts; they are estimates and should not drive refunds or escalations on their own.

Botcadence voice agents target sub-second response times with natural barge-in, and voice language coverage depends on the voice model chosen for the agent; see voice agents and the entry on TTS.