Summary

  • RFC 741 treated packet voice as a call to coordinate, not simply a stream of samples: two hosts had to agree on addressing, encoding, timing and readiness before speech data began.
  • Its four directional control and data roles, optional delay probe and human answer state show what the early ARPANET voice design actually specified; the RFC's report of four sites does not prove broad deployment or modern VoIP.

Analysis

A voice call needed more than a data protocol

The pause before a conversation was part of the protocol. One computer had to locate the other, discover whether their voice equipment could work together, and hold off on sending speech until the receiving side was ready. A reliable file transfer could try again after a missing message. A real-time voice system had a different problem: repeating late speech would not restore the moment in which it belonged.

Danny Cohen's RFC 741 made that difference explicit. Its cover is dated 22 November 1977, while its title page identifies the text as NSC Note 68, dated 29 January 1976 and revising three earlier notes. The Network Secure Communications project sought to demonstrate low-bandwidth, high-quality, full-duplex digital voice over packet-switched networks. The preface says that existing encryption devices could secure the digitized speech; NVP was one part of the effort, not the encryption mechanism itself.

The acknowledgments give the document a practical foothold. Cohen wrote that NVP was first implemented in December 1973 and had been used for local and transnet real-time voice over the ARPANET. He named the Information Sciences Institute, Lincoln Laboratory, Culler-Harrison and Stanford Research Institute, and listed different host and vocoder combinations. That is valuable evidence of cross-site work. It is not a census of users, a measure of call quality or proof that the protocol became a general telephone service.

Four directional roles made a call legible

RFC 741 begins from a mismatch: the ARPANET host-to-host protocol had been optimized for data transfer and had proved unsuitable for interactive voice. NVP separated control messages from voice data and gave the two sides distinct jobs to coordinate. In its terminology, the upper eight bits of a 12-bit MESSAGE-ID were the LINK; the lower four were the SUB-LINK. These were logical message identifiers, not four physical wires or modern IP ports.

For one voice communication, the RFC assigned four directional roles. L carried control from caller to answerer, while K carried control back. Speech data went caller-to-answerer on L+1 and answerer-to-caller on K+1. The values for L and K came from an octal range, 340 through 375; they did not have to differ. Initial contact used link 377.

The first message on 377 named the caller and called party and proposed K. The answerer could refuse, or accept and assign L. The caller then called again on L, and the answerer began a compatibility negotiation. One side proposed WHAT parameters and HOW choices; the other accepted or rejected them. The negotiation could cover vocoding, sample period, version, maximum message length and parcel size. The RFC defined LPC and CVSD options, among others, so a common call did not require identical equipment at both ends.

Negotiation did not mean the person had answered

After the initial negotiation succeeded, the answerer rang a bell and sent a RINGING message. Only after a human answered did READY indicate that data could flow. The specification allowed READY without a preceding RINGING, but it still distinguished a compatible connection from a person being ready to speak.

The protocol also left room for the conversation to change. Either party could request renegotiation after the link values had been assigned. Those links remained fixed, but previously negotiated settings had to be negotiated again; speech messages were ignored until READY confirmed success. An optional ECHO REQUEST could help measure delay. The RFC said that echo support was not compulsory and that a missing echo must not end the connection. It was a measurement hook, not a latency guarantee.

That boundary is easy to lose when the document is read as an early VoIP standard. NVP's caller and callee fields encoded host, IMP and extension information. They addressed communication units; they did not authenticate a human being or decide who was authorized to speak. The project's secure-voice ambition depended on existing encryption devices outside this call-control protocol. RFC 741 described a way to establish and coordinate a voice exchange, not a complete security or telephone service.

RFC 980's 1986 protocol-ordering guide later placed NVP-II among “Minor Host Protocols.” That label tells us where the catalog filed it. It does not show how many networks still ran it. The durable historical record is narrower and more useful: by the mid-1970s, ARPANET researchers had implemented voice across several sites and had specified the negotiations, message identifiers and human readiness needed to make unlike machines converse. The network carried parcels only after the hosts had agreed what those parcels meant.

Sources