September 2026 Summaries

Deepgram ranked 17th on RepVue’s September 2026 list of the Best Venture-Backed Companies to Sell For, a ranking based on anonymous, verified sales-professional ratings rather than company applications or nominations. RepVue evaluates employers across compensation, quota attainment, leadership, product-market fit, lead flow, culture, diversity, and professional development, with the venture-backed category aimed at candidates considering high-growth companies with potential equity upside. A cited Deepgram account executive praised the company’s functioning and improving product, particularly its speech-to-text capabilities, and identified text-to-speech as a potentially significant growth opportunity within the expanding voice AI market. The piece argues that rep sentiment can help job seekers assess questions about quotas, leadership, product quality, and equity prospects, while noting that such rankings cannot answer every consideration; Deepgram also states that it is expanding its account executive team.

A Voice Comic demo built on Deepgram’s Voice Agent API illustrates how mid-call changes to speech recognition models, language models, and voices can succeed server-side even when expected WebSocket acknowledgements such as ListenUpdated or ThinkUpdated fail to reach the client. To prevent the interface from remaining stuck during model changes, the implementation treats acknowledgements as a fast path and uses a three-second timeout that silently reconnects with fresh Settings, no greeting, and a resume identifier to preserve session-related state. The author found that UpdateListen acknowledgements were intermittently missing, especially when switching between Flux STT v2 and Nova-3 v1, while UpdateThink acknowledgements could disappear when sent immediately alongside UpdateListen, leading to a recommendation to sequence LLM changes separately or include both settings in a reconnect. The demo also distinguishes idle-stream behavior by keeping the agent and Nova-3 streams alive with periodic KeepAlive messages while closing and reopening Flux STT streams, whose v2 endpoint does not support KeepAlive; because Flux turn indexes restart on each new socket, the UI combines a socket epoch with the turn index to avoid overwriting prior transcript panels. SpeakUpdated acknowledgements were consistently received in the author’s tests, so voice changes do not use the fallback timer, and the project includes a public repository, hosted demo, event log, and live tests for reproducing the behavior.

Deepgram has introduced Nova-3 Pharma, a speech-to-text model designed specifically for pharmaceutical workflows in which accurately recognizing medication names is especially important for applications such as pharmacy IVRs, refill automation, contact-center agents, and clinical documentation. Built on the company’s medical speech foundation and trained on more than 5,000 drug names, the model aims to distinguish uncommon and phonetically similar pharmaceutical terms while retaining general transcription quality. In Deepgram’s benchmark using English medical audio from settings including telehealth, clinical dictation, and triage, Nova-3 Pharma ranked first among evaluated models for drug-name Keyword Recognition Rate, reporting 91.59% for batch transcription and 91.28% for streaming, and for lowest overall Word Error Rate, reporting 10.17% in batch and 11.69% in streaming. It is available through Deepgram’s hosted API and self-hosted deployments, supports batch and streaming transcription, and works across eight English language variants.

Deepgram has made its India endpoint generally available, allowing customers to run speech-to-text, text-to-speech, Voice Agent, and text-intelligence workloads with audio storage and inference hosted in AWS’s Hyderabad region. The service is intended particularly for organizations in regulated sectors such as banking, payments, and insurance that need to keep customer voice data within India, avoiding the cross-border exposure of other regional endpoints or the operational burden of self-hosting. Customers can use the same APIs, SDKs, API keys, models, and launch pricing as Deepgram’s Global, EU, and Australia endpoints by changing their base URL to api.in.deepgram.com. By default, audio, transcripts, and speech output are stored and processed in India, while opting out of model improvement keeps customer content out of training-related workflows; however, billing and operational metadata remain processed in the United States, and managed third-party LLMs used by Voice Agent may process requests outside India. Deepgram positions the fully managed endpoint, alongside its SOC 2 certifications, as part of a wider regional expansion strategy aimed at providing local voice AI processing without sacrificing model capabilities.

Live voice agents must manage emotional TTS without the manual review available in traditional voiceover workflows, using either inline markup generated by the LLM, session-level synthesis parameters, or cross-turn model state. Using Deepgram’s Flux TTS as an example, the text explains that inline markup enables turn-specific expression but adds token cost, buffering requirements, and risks from malformed tags, while Flux’s connection-level expressivity setting ranges from calm to animated and provides consistent delivery without added token or latency costs but remains fixed for a WebSocket session and is labeled beta. Flux also preserves prosodic context across turns within an active connection, although this state cannot be directly adjusted and resets only when a new connection is opened. Because emphasis involves linked changes in pitch, duration, and energy, emotion settings can affect voices differently and cannot reliably solve pronunciation issues, which may instead require tools such as Aura-2’s IPA lexicon controls. The guidance recommends default expressivity and voice selection for accuracy-sensitive healthcare, financial, and IVR readbacks, while allowing carefully auditioned non-default settings or markup for lower-risk conversational, sales, and reminder use cases. It advises testing exact voices, scripts, number-heavy content, reconnection behavior, latency, and model updates before deployment, since stronger expressivity settings can increase repetitions, omitted words, added words, and pronunciation errors.

Voice-agent function calling enables automated responses to requests such as order-status inquiries, but telephone interactions impose strict latency, speech-recognition, and conversational-design requirements because callers experience backend delays as silence and spoken identifiers can be misrecognized. A recommended order-status flow uses three functions—an obligatory filler message, customer identification, and order retrieval—with chained lookups requiring identity resolution before order data is accessed. Sensitive database and third-party calls should generally execute server-side so credentials remain protected, while backend authorization must independently verify that returned customer records belong to the caller rather than relying on the language model or function schema. The guide emphasizes defining natural spoken responses in advance for timeouts, empty results, and incorrect-record risks, setting explicit lookup timeouts, and returning structured, speakable responses instead of allowing indefinite waits or technical errors. It also advises measuring end-to-end latency and backend P95 performance, caching authentication keys, logging function-call metadata and non-audio events for observability, testing successful, missing-record, and delayed-handler scenarios before launch, and starting with a limited toolset because function-selection accuracy declines as tools proliferate.

Voice AI telephony integrations generally follow either a CPaaS media-streaming model, where platforms such as Twilio or Vonage retain call control and send audio over WebSockets, or an SBC/SIP hand-off model, where existing PBX or contact-center systems retain telephony control while an SBC such as AudioCodes VoiceAI Connect connects speech services like Deepgram. The choice affects media routing, fallback design, deployment complexity, and the ability to preserve existing call infrastructure. Production planning also requires detailed review of contractual uptime terms, which vary in measurement methods, exclusions, claims processes, and credit remedies, alongside independent carrier SLA coverage. Resilience depends on both SIP trunk failover and regional speech-service endpoints, with regional concurrency limits and latency monitoring affecting capacity planning. International deployment adds country-specific number eligibility, carrier-latency testing, data residency requirements, and potential use of private cloud or self-hosting. The recommended approach is to begin with a one-way transcription pilot on an existing trunk, validate real-call quality and regional performance, then expand to bidirectional voice interactions after confirming regulatory, availability, and provider-failover requirements.

Pronunciation Error Rate (PER) is presented as a phoneme-level metric for evaluating text-to-speech accuracy in areas where broad measures such as Mean Opinion Score (MOS) and Word Error Rate (WER) can conceal costly errors involving medical terms, brand names, financial language, and alphanumeric identifiers. PER calculates phoneme substitutions, deletions, and insertions against a reference pronunciation, enabling teams to identify specific sound-level failures, while forced alignment is described as more reliable for precise evaluation than ASR round-trip testing, which may generate substantial false alarms. Production implementations combine custom pronunciation dictionaries, SSML phoneme controls, automated WER and Keyword Error Rate monitoring, periodic MOS testing, and business-specific thresholds, such as under 5% KER for critical healthcare terminology or above 90% accuracy for contact-center vocabulary. The approach also emphasizes testing providers with domain-specific corpora, monitoring pronunciation alongside latency, setting alerts for degradation, and using captured failures to refine dictionaries and validate fixes before deployment.

Deepgram has appointed Ed Anuff as Chief Product Officer, bringing more than 30 years of experience in data, APIs, enterprise platforms, and product leadership at organizations including IBM, DataStax, Apigee, and Wired. Anuff joined after advising Deepgram since 2021, citing the company’s evolution from speech-to-text technology toward a production-scale voice AI platform and his belief that voice will become AI’s primary interaction interface within two years. He argues that voice offers richer contextual signals than typed prompts, including pauses, pace, confidence, and corrections, while enabling AI interactions away from screens. Anuff says the industry is shifting from isolated voice AI demonstrations to reliable systems capable of handling real-world complexity, requiring low latency, accuracy across languages and noisy environments, interruption recovery, testing, observability, and robust infrastructure. His priorities include improving Deepgram’s natural and emotionally resonant text-to-speech capabilities, connecting voice agents to actions and transactions, developing an open-source strategy, and helping developers and enterprises more easily deploy scalable, context-aware voice systems.

Effective voice-agent testing evaluates complete conversations rather than relying solely on speech-to-text metrics such as word error rate, using a starter suite of nine realistic scenarios that cover self-corrections, interruptions, silence, accents, emotional speech, noise, out-of-scope requests, and wrong numbers. Each recorded call should be assessed against a clearly written end state using measures including task completion, turns to resolution, appropriate escalation, and false statements, with system records and logs supporting automation while human reviewers judge nuanced behavior. Synthetic TTS clips are useful for early testing but cannot fully replicate real callers’ timing, overlap, disfluencies, emotional delivery, or telephony conditions, so testing should progress from uploaded audio to real calls over the deployed transport and then to production monitoring. Because small prompt or model changes can cause unpredictable whole-call regressions, teams should version prompts, change one element at a time, and rerun the entire suite after every change. After launch, monitoring should focus on latency, language and transport differences, transfer success, silence-triggered failures, and unexpected intents, with newly observed failures added back into the scenario suite to continually improve coverage.

Prosody support for multilingual voice agents depends on separate transcription, conversational recognition with turn detection, and speech-synthesis layers, represented by Nova-3, Flux Multilingual, and Aura-2 respectively, so language availability must overlap across all required layers for full agent deployment. Nova-3 supports transcription of Mandarin, Cantonese, and Thai, but Flux Multilingual’s ten-language conversational coverage excludes tonal languages, limiting those markets to transcription workloads rather than live conversations; Aura-2 provides multilingual synthesis, including Japanese, while Flux TTS is English-only. The distinction is important because tone languages use pitch to distinguish word meanings, pitch-accent languages such as Japanese use accent position to shape word melody, and other languages employ pitch, phrasing, duration, loudness, and timing differently for emphasis and politeness. Real-time recognition is especially difficult in tonal languages because pitch must simultaneously convey lexical meaning and help identify turn endings, while streaming systems may need additional audio context before accurately resolving a tone. The guide recommends evaluating support separately for each layer, testing difficult real-world audio and turn behavior, using native listeners to assess synthesized politeness and register, and avoiding production deployment in languages that do not meet acceptance criteria across transcription, conversation, and synthesis.

Voice AI latency strongly affects how callers judge an agent’s confidence, competence, and attentiveness, because people expect conversational turn transitions to occur in roughly a quarter of a second and may interpret delays near or above one second as hesitation or a dead line. The most meaningful measure is the p90 silence from the end of a user’s speech to the agent’s first audible response, rather than full-response time or median latency, since slow outlier calls shape user trust and abandonment. Systems must balance fast end-of-turn detection against premature interruptions, which can cut off speech and require costly repeat turns, while downstream LLM and text-to-speech processing add further delay. Appropriate latency budgets vary by context: phone IVR systems need especially fast responses, in-app assistants can provide visible acknowledgment but still require responsive audio, drive-through systems must prioritize recognition accuracy amid noise, and live captions must balance synchronization with accuracy. The guidance recommends measuring performance per turn under real production conditions, setting targets by deployment and interaction type, streaming responses where possible, testing with diverse speech patterns, and using fillers or status cues only to smooth residual delays rather than conceal slow processing.

Deepgram’s `filler_words` setting affects seven detected speech tokens but, by default, documentation indicates that it removes only “uh” and “um,” while at least “uh-huh” can remain in transcripts and the behavior of other backchannel sounds is unclear. This distinction matters because “uh” and “um” generally signal hesitation, whereas tokens such as “uh-huh,” “uh-uh,” “mhmm,” and “mm-mm” may communicate agreement, disagreement, or listener attention, potentially altering the meaning of legal records, research data, clinical interactions, meeting summaries, and language-model inputs. The recommended setting depends on the use case: verbatim-sensitive contexts such as court reporting, conversation research, and speaking coaching may require retaining tokens, while clinical documentation, captions, and recaps often favor cleaner output. Because Deepgram’s documentation contains ambiguities and conflicting statements about model support, language coverage, and retained backchannels, developers are advised to test their exact model and configuration, preserve raw transcripts where possible, and explicitly process any remaining non-lexical tokens downstream.

Paralinguistic cues are vocal signals that convey meaning beyond words, including prosody, voice quality, disfluency, timing and silence, and non-speech vocalizations such as laughter or sighs. Production transcription systems currently support timing measurements through word timestamps and speaker diarization, while some providers can retain filler words like “uh” and “um,” but they generally do not label repairs, false starts, or other disfluencies without post-processing. Prosody and voice-quality features, including pitch, loudness, breathiness, jitter, shimmer, and harmonics-to-noise ratio, can be measured from raw audio using tools such as openSMILE and Praat, but are not broadly returned by transcription APIs and lack consistent mappings from acoustic values to perceptual labels. Non-speech vocalizations have the weakest commercial support, with AssemblyAI documenting laughter tags but no major provider offering documented sigh or crying detection. The guidance recommends preserving audio alongside timestamped, diarized transcripts, since deleting recordings permanently prevents later extraction of prosodic and voice-quality information.

End-of-turn detection determines when a voice-agent caller has finished speaking and must balance false interruptions against response latency. Fixed silence timers and voice activity detectors infer completion from quiet, but timers confuse mid-sentence pauses with finished turns while VADs cannot distinguish hesitation from completion and may be affected by noise; both require configuration choices that can add delay or errors. Models that fuse acoustic cues such as prosody and timing with transcript-level semantic context can better recognize whether a thought is complete, though they introduce compute costs and confidence-threshold tuning, with a silence timeout retained as a fallback. Because no universal standard defines an acceptable false-interruption rate, platforms should establish their own baseline through stratified sampling of deployed calls, define interruptions consistently, compare controlled configuration changes on the same sample, and examine results by language, customer, and queue. Eager end-of-turn approaches can begin LLM generation 150–250 milliseconds sooner on likely completions, but increase discarded drafts and LLM calls when callers continue speaking, making them more suitable for inexpensive, short transactional interactions than dictation or costly workloads. The recommended approach is to begin with default settings, measure interruption rates and response times before tuning, test multilingual and noisy conditions, and track eager-generation waste where enabled.

Deepgram’s Flux TTS expressivity setting is presented as a session-level control for delivery register, adjusting pitch variation and pacing on an integer scale from -2 for calmer speech to +2 for more animated speech, while 0 remains each voice’s production-tuned default. The guide argues that voice agents often sound “robotic” because their delivery remains in one register throughout a conversation, making configuration testing a less costly first step than changing models. Expressivity applies only to Flux TTS, not Aura-2, persists automatically across turns, cannot be changed during an active connection, and differs from emotional TTS systems that require per-line labels. Although Flux TTS is generally available, expressivity is beta, with non-default values carrying higher risks of hallucinations and pronunciation errors; therefore, the default is recommended for production and regulated or high-stakes uses, while calmer settings may suit support, IVR, and de-escalation, and more animated settings may fit consumer or outbound applications. Users are advised to test real production scripts on their selected voice and setting, handle invalid-parameter connection failures, and repeat validation after model updates, as voices vary in their response to the same value.

Voice agents using energy-based voice activity detection often mistake listener backchannels such as “uh-huh” and “mhmm” for interruptions, causing unnecessary stops, while failing to distinguish them from genuine corrections or attempts to take the floor. The proposed approach uses transcript-aware turn detection, filler-word recognition, token classification, and a short post-token wait window: standalone affirmative backchannels should generally let the agent continue, negative forms such as “uh-uh,” “nuh-uh,” and “mm-mm” should prompt an immediate yield, and affirmative tokens followed by additional speech should also be treated as a turn bid. Because ambiguous words like “yeah,” “right,” and “okay” depend heavily on prosody that may not be available in speech-to-text outputs, the guidance recommends relying on token identity, timestamps, continued speech, and contextual turn-detection models such as Flux STT rather than simple lookup rules alone. It also advises logging each interruption event, transcript, policy decision, and subsequent caller behavior to measure false-interruption rates and missed corrections, establish a baseline before tuning, and adjust policies according to the cost of errors in different conversational stages, languages, and use cases.