Requirements for Distributed Control of ASR, SI/SV and TTS Resources

There are multiple IETF protocols for establishment and termination of media sessions (SIP), low-level media control (MGCP and MEGACO), and media record and playback (RTSP). This document focuses on requirements for one or more protocols to support the control of network elements that perform Automated Speech Recognition (ASR), speaker identification or verification (SI/SV), and rendering text into audio, a.k.a. Text-to-Speech (TTS). Many multimedia applications can benefit from having automatic speech recognition (ASR) and text-to-speech (TTS) processing available as a distributed, network resource. This requirements document limits its focus on the distributed control of ASR, SI/SV and TTS servers. There are a broad range of systems which can benefit from a unified approach to control of TTS, ASR, and SI/SV. These include environments such as VoIP gateways to the PSTN, IP Telephones, and wireless mobile devices who obtain speech services via servers on the network. To date, there are a number of proprietary ASR and TTS API's, as well as two IETF drafts that address this problem , . However, there are serious deficiencies to the existing drafts. In particular, they mix the semantics of existing protocols yet are lose enough to other protocols as to be confusing to the implementer. This document sets forth requirements for protocols to support distributed speech processing of audio streams. For simplicity, and to remove confusion with existing protocol proposals, this document presents the requirements as being for a "new protocol" that addresses the distributed control of speech resources It refers to such a protocol as "SPEECHSC", for Speech Services Control Protocol.

The following is the SPEECHSC framework for speech processing.

+-------------+ | Application | | Server |\ +-------------+ \ SPEECHSC SIP, VoiceXML, / \ etc. / \ +------------+ / \ +-------------+ | Media |/ SPEECHSC \---| ASR, SI/SV | | Processing |-------------------------| and/or TTS | RTP | Entity | RTP | Server | =====| |=========================| | +------------+ +-------------+ The "Media Processing Entity" is a network element that processes media. It may be either a pure media handler, or also have an associated SIP user agent, VoiceXML browser or other control entity. The "ASR, SI/SV and/or TTS Server" is a network element which performs the back-end speech processing. It may generate an RTP stream as output based on text input (TTS) or return recognition results in response to an RTP stream as input (ASR, SI/SV). The "Application Server" is a network element that instructs the Media Processing Entity on what transformations to make to the media stream. Those instructions may be established via a session protocol such as SIP, or provided via a client/server exchange such as VoiceXML. The framework allows either the Media Processing Entity or the Application Server to control the ASR or TTS Server using SPEECHSC as a control protocol, which accounts for the speechsc protocol appearing twice in the diagram. Physical embodiments of the entities can reside in one physical instance per entity, or some combination of entities. For example, a VoiceXML Gateway may combine the ASR and TTS functions on the same platform as the Media Processing Entity. Note that VoiceXML Gateways themselves are outside the scope of this protocol. Likewise, one can combine the Application Server and Media Processing Entity, as would be the case in an interactive voice response (IVR) platform. One can also decompose the Media Processing Entity into an entity that controls media endpoints and entities that process media directly. Such would be the case with a decomposed gateway using MGCP or megaco. However, this decomposition is again orthogonal to the scope of SPEECHSC.

To the extent feasible, the SPEECHSC framework SHOULD use existing protocols.

In meeting the requirement of , the SPEECHSC framework MUST NOT redefine the semantics of an existing protocol. Said differently, we will not break existing protocols or cause backward compatibility problems.

To the extent feasible, SPEECHSC SHOULD NOT duplicate the functionality of existing protocols. For example, network announcements using SIP and RTSP already define how to request playback of audio. The focus of SPEECHSC is new functionality not addressed by existing protocols or extending existing protocols within the strictures of the requirement in . Where an existing protocol can be gracefully extended to support SPEECHSC requirements, such extensions are acceptable alternatives for meeting the requirements. As a corollary to this, the SPEECHSC should not require a separate protocol to perform functions that could be easily added into the SPEECHSC protocol (like redirecting media streams, or discovering capabilities), unless it is similarly easy to embed that protocol directly into the SPEECHSC framework.

The SPEECHSC framework SHOULD employ protocol elements known to result in efficient operation. Techniques to be considered include: Re-use of transport connections across sessions Piggybacking of responses on requests in the reverse direction Caching of state across requests

The SPEECHSC framework MUST be compliant with the IAB OPES framework. The applicability of the SPEECHSC protocol will therefore be specified as occurring between clients and servers at least one of which is operating directly on behalf of the user requesting the service.

To the extent feasible, the SPEECHSC framework SHOULD exploit existing schemes for supporting service location and load balancing, such as the Service Location Protocol or DNS SRV records. Where such facilities are not deemed adequate, the SPEECHSC framework MAY define additional load balancing techniques.

The SPEECHSC framework MUST permit multiple services to operate on a single media stream so that either the same or different servers may be performing speech recognition, speaker identification or verification, etc. in parallel.

The SPEECHSC framework MUST allow a 1:N mapping between session and RTP channels. For example, a single session may include an outbound RTP channel for TTS, an inbound for ASR and a different inbound for SI/SV (e.g. if processed by different elements on the Media Resource Element). Note: All of these can be described via SDP, so if SDP is utilized for media channel description, this requirement is met "for free".

The SPEECHSC framework must have sufficient capabilities to address the critical needs of handicapped users. In particular, the set of requirements set forth in RFC3351 MUST be taken into account by the framework.

The SPEECHSC framework MUST allow a Media Processing Entity or Application Server, using a control protocol, to request the TTS Server to playback text as voice in an RTP stream.

The SPEECHSC framework MAY assume that all TTS servers are capable of reading plain text. For reading plain text, framework MUST allow the language and voicing to be indicated via session parameters. For finer control over such properties, see .

The SPEECHSC framework MUST support SSML[3] <speak> basics, and SHOULD support other SSML tags. The framework assumes all TTS servers are capable of reading SSML formatted text.

The Speechsc framework assumes all TTS servers accept text over the SPEECHSC connection for reading over the RTP connection. The framework assumes the server can accept text either "by value" (embedded in the protocol), or "by reference" (e.g. by de-referencing a URI embedded in the protocol).

The SPEECHSC framework MUST be capable of explicitly indicating the document type of the text to be processed, as opposed to forcing the server to infer the content by other means.

The SPEECHSC framework MUST be capable of establishing the control channel between the client and server on a per-session basis, where a session is loosely defined to be associated with a single "call" or "dialog". The protocol SHOULD be capable of maintaining a long-lived control channel for multiple sessions serially, and MAY be capable of shorter time horizons as well, including as short as for the processing of a single utterance.

The SPEECHSC framework MUST NOT require the controlling element (application server, media processing entity) to accept or originate media streams. Media streams MAY source & sink from the controlled element (ASR, TTS, etc.).

The Speechsc framework MUST support "VCR controls", and MUST allow for servers with varying capabilities to accommodate such controls. These capabilities include: The ability to jump in time to the location of a specific marker. The ability to jump in time, forwards or backwards, by a specified amount of time. Valid time units MUST include seconds, words, paragraphs, sentences, and markers. The ability to increase and decrease playout speed. The ability to fast-forward and fast-rewind the audio, where snippets of audio are played as the server moves forwards or backwards in time. The ability to pause and resume playout. The ability to increase and decrease playout volume.

The SPEECHSC framework must support the specification of session parameters, such as language, prosody and voicing.

The SPEECHSC framework MUST accommodate speech markers, with capability at least as flexible as that provided in SSML. The framework MUST further provide an efficient mechanism for reporting that a marker has been reached during playout.

The SPEECHSC framework MUST allow a Media Processing Entity or Application Server to request the ASR Server to perform automatic speech recognition on an RTP stream, returning the results over SPEECHSC.

The Speechsc framework assumes that all ASR servers support thh VoiceXML speech recognition grammar specification (SRGS) for speech recognition .

The Speechsc framework assumes all ASR servers are capable of accepting grammar specifications either "by value" (embedded in the protocol), or "by reference" (e.g. by de-referencing a URI embedded in the protocol). The latter MUST allow the indication of a grammar already known to, or otherwise "built in" to the server. The framework and protocol further SHOULD exploit the ability to store and later retrieve by reference large grammars which were originally supplied by the client.

The SPEECHSC framework protocol MUST be able to explicitly convey the grammar format in which the grammar is encoded and MUST be extensible to allow for conveying new grammar formats as they are defined.

The Speechsc framework SHOULD exploit sharing grammars across sessions for servers which are capable of doing so. This supports applications with large grammars for which it is unrealistic to dynamically load. An example is a city-country grammar for a weather service.

The SPEECHSC framework MUST accommodate at a minimum all of the protocol parameters currently defined in MRCP In addition there SHOULD be a capability to reset parameters within a session.

The SPEECHSC framework MUST support a method directing the ASR Server to capture the input media stream for later analysis and tuning of the ASR engine.

The SPEECHSC framework MUST allow a Media Processing Entity to request the SI/SV Server to perform speaker identification or verification on an RTP stream, returning the results over SPEECHSC.

The SPEECHSC framework MUST accommodate an identifier for each verification resource and permit control of that resource by ID, because voiceprint format and contents are vendor specific.

The SPEECHSC framework MUST work with SI/SV servers which maintain state to handle multi-utterance verification.

The SPEECHSC framework MUST support a method for capturing the input media stream for later analysis and tuning of the SI/SV engine. The framework may assume all servers are capable of doing so. In addition the framework assumes that the captured stream contains enough timestamp context (e.g. the NTP time range from the RTCP packets which corresponds to the RTP timestamps of the captured input) to ascertain after the fact exactly when the verification was requested.

The SPEECHSC framework SHOULD be extensible to additional functions associated with SI/SV, such as prompting, utterance verification, and retraining.

One very important requirement for an interactive speech-driven system is that user perception of the quality of the interaction depends strongly on the ability of the user to interrupt a prompt or rendered TTS with speech. Interrupting, or barging, the speech output requires more than energy detection from the user's direction. Many advanced systems halt the media towards the user by employing the ASR engine to decide if an utterance is likely to be real speech, as opposed to a cough, for example.

To achieve low latency between utterance detection and halting of playback, many implementations combine the speaking and ASR functions. The SPEECHSC framework MUST support such full-duplex implementations.

Good spoken user interfaces typically depend upon the ease with which the user can accomplish his or her task. When making use of Speaker Identification or Verification technologies, user interface improvements often come from the combination of the different technologies: simultaneous identity claim and verification (on the same utterance), simultaneous knowledge and voice verification (using ASR and verification simultaneously). Using ASR and verification on the same utterance is in fact the only way to support rolling or dynamically-generated challenge phrases (e.g., "say 51723"). The SPEECHSC framework MUST support such parallel service implementations.

It is optionally of interest that the SPEECHSC framework support more complex remote combination and controls of speech engines: Combination in series of engines that may then act on the input or output of ASR, TTS or Speaker recognition engines. The control MAY then extend beyond such engines to include other audio input and output processing and natural language processing. Intermediate exchanges and coordination between engines Remote specification of flows between engines. These capabilities MAY benefit from service discovery mechanisms (e.g. engines, properties and states discovery).

The framework assumes that SDP will be used to describe media sessions and streams. The framework further assumes RTP carriage of media, however since SDP can be used to describe other media transport schemes (e.g. ATM) these could be used if they provide the necessary elements (e.g. explicit timestamps). The working group will not be defining distributed speech recognition methods (DSR), as exemplified by the ETSI Aurora project. The working group will not be recreating functionality available in other protocols, such as SIP or SDP. TTS looks very much like playing back a file. Extending RTSP looks promising for when one requires VCR controls or markers in the text to be spoken. When one does not require VCR controls, SIP in a framework such as Network Announcements works directly without modification. ASR has an entirely different set of characteristics. For barge-in support, ASR requires real-time return of intermediate results. Barring the discovery of a good reuse model for an existing protocol, this will most likely become the focus of SPEECHSC.

Protocols relating to speech processing must take security into account. This is particularly important as popular uses for TTS include reading financial information. Likewise, popular uses for ASR include executing financial transactions and shopping. We envision that rather than providing application-specific security mechanisms in SPEECHSC itself, the resulting protocol will employ security machinery of either containing protocols or the transport on which it runs. For example, we will consider solutions such as using TLS for securing the control channel, and SRTP for securing the media channel. Third-part dependencies necessitating transitive trust will be minimized or explicitly dealt with through the authentication and authorization aspects of the protocol design. In addition to the security machinery needed by the protocol itself, there are considerations for the implementation and deployment of the clients and servers themselves. For example, speaker verification and identification employs voiceprints whose privacy and integrity must be maintained. While strictly speaking out of scope of the protocol itself, such considerations will be carefully considered and accommodated during protocol design, and will be called out as part of the applicability statement accompanying the protocol specification(s).