Stretto Rebuilds Bankruptcy Intake with Amazon Nova 2 Sonic
Summary:Stretto, the AWS Generative AI Innovation Center, and CI&T deployed Tara, a voice agent for multi-stage bankruptcy intake. Unlike a conventional ASR-to-LLM-to-TTS pipeline, Tara uses Amazon Nova 2 Sonic through Amazon Bedrock as a native speech-to-speech model, while Twilio streams caller audio to an ECS/Fargate voice bridge. Claude 3 Haiku extracts typed fields asynchronously after the call, while deterministic attorney-defined rules—not the model—determine lead qualification. The most reusabl
Stretto has replaced the Bankruptcy.com web-form intake with Tara, a voice agent for bankruptcy lead qualification. Amazon Nova 2 Sonic handles native speech understanding and generation during the live call, Claude 3 Haiku extracts typed fields afterward, and attorney-defined rules plus human review determine the final qualification outcome. The important contribution is not merely replacing a form with a conversation; it is a production architecture designed around streaming audio, long-session continuity, asynchronous persistence, and explicit responsibility boundaries in a sensitive financial workflow.
Architecture: Separating live interaction from offline structuring
Calls enter through Twilio, where Media Streams carries audio over a WebSocket. Traffic passes through AWS WAF and an Application Load Balancer before reaching a voice bridge running on Amazon ECS on Fargate. The bridge runs continuously across three Availability Zones and opens a bidirectional audio stream with Amazon Nova 2 Sonic through Amazon Bedrock.
After the call, the transcript enters an asynchronous processor through Amazon SQS and a dead-letter queue. The processor invokes Claude 3 Haiku on Bedrock to convert names, dollar amounts, addresses, vehicles, and other information into typed fields, then writes the results to Amazon Aurora Serverless v2 in a Multi-AZ configuration. Attorney-defined rules classify the lead as qualified, disqualified, or requiring more information, while attorneys retain final authority.
This split has two operational benefits. The real-time path is limited to conversation and tool calls, preventing structured extraction from dominating time-to-first-audio. The asynchronous path can scale and retry independently, so an extraction failure does not immediately terminate an active call. It also creates an eventual-consistency window, requiring idempotency keys, queue-backlog alerts, and dead-letter replay to handle duplicate or late events safely.
What native speech-to-speech changes
A conventional ASR-LLM-TTS pipeline processes audio chunking, transcription, text inference, and speech synthesis in sequence. Its time-to-first-audio can be viewed conceptually as the sum of capture buffering, ASR time, first-token latency, TTS time, and playback buffering. Jitter from any stage accumulates, while the explicit text boundary can discard prosody, hesitation, and other paralinguistic cues.
Nova 2 Sonic accepts speech and produces speech within one foundation model, removing several explicit handoffs and intermediate representations. Stretto says its previous solution exceeded 500 milliseconds to first audio, while Nova 2 Sonic was designed for lower-latency dialogue. The agent was also expected to adjust its pace and tone for stressed, hesitant, or frustrated callers.
End-to-end responsiveness, however, is not determined by model inference alone. Twilio capture and encoding, WebSocket jitter, Bedrock queueing, container scheduling, output buffering, and playback can dominate the critical path. The post does not disclose Nova 2 Sonic's actual TTFA distribution, so the 500-millisecond figure is only a legacy baseline rather than proof of the replacement's latency. Production acceptance should report p50, p95, and p99 TTFA under packet loss, cross-AZ traffic, and elevated concurrency.
Barge-in is more than a stop button. Effective interruption handling requires detecting user speech, cancelling in-flight generation, flushing queued audio, preserving conversational state, and resuming from the correct point. Because the article does not expose the model's internal VAD, cancellation, and buffering mechanisms, teams should inject interruptions during disclosures, in the middle of questions, and while the agent is speaking.
Continuing interviews beyond the session limit
A complete bankruptcy interview may run for 15–20 minutes, exceeding the duration limit of one Nova 2 Sonic streaming session. Stretto's voice bridge tracks the lifetime of each bidirectional stream. Before reaching the limit, it opens a new stream with the same system prompt and tool configuration, supplies the conversation so far as text history, waits for the current response to finish, switches the caller's audio to the new stream, and closes the old one.
This avoids a hard reset at the session boundary, but introduces three risks. Text history cannot fully preserve prosody, emotional cues, and speech timing, so the new stream may not maintain identical delivery. Longer conversations also require more tokens to reconstruct context, increasing input and inference costs. Finally, new-stream initialization and warm-up must be tightly scheduled; even a response-boundary switch can produce an audible pause if the handoff is not coordinated.
A stronger implementation separates canonical facts from conversational history. The real-time model collects information, while the application stores confirmed fields, tool results, and version metadata as structured records. A continued session then receives stable fact IDs, a compact summary, and only the original text that is actually required.
Keeping consequential decisions outside the model
Tara can conduct the interview, invoke tools, and extract information, but it does not decide whether a caller qualifies for legal services. Qualification rules are defined by attorneys and executed after the call, with attorneys retaining final authority. This boundary is more robust than relying on prompt instructions alone because explicit rules can be audited, replayed, and versioned.
The test program covered happy paths, contradictory answers, mid-call corrections, refusals, attempts to manipulate the outcome, consent collection, disclosures, and interruptions at different points. Consent and disclosure delivery were logged, while model safety controls, deterministic business rules, and human review formed layered safeguards.
AWS states that Bedrock does not use prompts or model outputs to train models, and IAM with PrivateLink can keep model traffic inside the AWS network. These are platform-provider data-protection claims, not certification of this particular application. Debt, asset, and employment information still requires data minimization, explicit retention periods, access auditing, deletion workflows, regional residency controls, and a defined human-escalation path. Nova 2 Sonic and Bedrock are also not available in every AWS Region.
Reported results and their limits
| Metric | Reported result | Important qualification | | --- | --- | --- | | Per-minute voice cost | Approximately 95% lower than the previous solution | Based on April 2026 on-demand pricing in us-east-1; the old solution, discounts, and total ECS, SQS, Aurora, and other costs are not disclosed | | Structured extraction accuracy | Approximately 96% | No sample count, annotation protocol, field-level versus record-level definition, confidence interval, or refusal-case analysis | | Average completed-call duration | Approximately eight minutes | Completed calls only, which may omit interruptions, disconnects, and other failure outcomes | | Lead conversion | 20% relative improvement over the web form | This is not a 20-percentage-point increase; rollout had not covered every district, and channel-intent differences may also contribute | | Voice quality | Adapted across calm, stressed, hesitant, and frustrated states | No MOS, blind human preference, emotion-recognition accuracy, or comparison with the previous system |
The numbers show that the system has progressed beyond a demonstration, but they do not constitute an independently reproducible benchmark. In particular, a 95% reduction in model price is not the same as a 95% reduction in total operating cost. A complete unit-economics model must include concurrent voice-bridge capacity, Claude extraction, storage, monitoring, retries, and human review.
Core engineering trade-offs
- Native speech-to-speech versus modular control: Nova 2 Sonic removes several pipeline boundaries and improves vocal continuity, but provides less direct control over internal ASR, audio chunking, and generation behavior. It also introduces vendor dependency and regional availability constraints.
- Low latency versus long-range consistency: Session continuation removes the maximum-duration wall, but converts some non-verbal information into text. The more sensitive the conversation, the more important hesitation, self-correction, and vocal tension become, increasing the need for an external state layer and cross-session regression tests.
- Automation versus accountability: Models can lower the cost of collecting information, but should not make consequential legal qualifications. Deterministic rules and attorney review add process cost while providing explainability and a clear responsibility boundary.
- Managed services versus total cost of ownership: Bedrock removes model-infrastructure operations, but lower token pricing does not eliminate system cost. High-concurrency deployments must account for continuously running Fargate capacity, peak scaling, Multi-AZ redundancy, queue backlog, and downstream storage.
- Emotional interaction versus evaluation difficulty: Better tone adaptation can improve user experience, but a successful demonstration cannot establish reliability. Blind listening tests, emotional-state strata, noise and accent coverage, and inter-annotator agreement are needed.
Deployment guidance for developers
- Set a quality gate before integration: Validate emotional delivery, noise, accents, clarification, and interruption behavior against explicit thresholds for TTFA, interruption recovery, and field accuracy.
- Externalize conversational state: Persist structured facts, tool results, and schema versions in a database or event store. Continue sessions with necessary context rather than reconstructing state from the full dialogue.
- Make the asynchronous path idempotent: Construct idempotency keys from call IDs and extraction versions; monitor SQS backlog, dead-letter volume, and replay outcomes to prevent duplicate writes or overwritten human corrections.
- Separate model output from business decisions: Run qualification, monetary thresholds, and legal routing through a versioned rules engine. Retain logs for model input, output, rule version, and human overrides.
- Design degraded operation: Provide retry, callback, voicemail, and human-handoff paths for Bedrock timeouts, voice-bridge failures, dropped long connections, and queue backlogs. Sensitive workflows should not depend on a single automatic retry.
- Calculate cost per qualified lead: Put complete-call cost, per-minute model cost, extraction cost, human review, and qualified-lead count into one equation instead of comparing model prices in isolation.
The reusable pattern in the Stretto deployment is explicit: a native speech model handles interaction, an asynchronous model handles structure, a rules engine handles qualification, and humans retain final responsibility. Whether such a system can safely enter finance or other high-sensitivity domains depends less on a polished demo than on tail latency, long-session continuity, auditability, and complete operating cost.
— END —Source: AWS Machine Learning Blog (2026-11-18)
Tags: #Amazon Nova 2 Sonic #Voice Agents #Speech-to-Speech #Amazon Bedrock #Enterprise AI
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments