Product teams building call analysis pipelines often tune their LLM prompts carefully and neglect the layer feeding the model its input. If the transcript arrives with dropped negations, misattributed utterances, or misheard entities, no prompt adjustment fixes it. The error is already upstream, and every CRM entry, sentiment score, and coaching record downstream inherits it.
Conversation intelligence (CI) transforms raw audio into structured data for sales coaching, support QA, and product roadmapping. Whether you buy an off-the-shelf CI platform or build your own pipeline, the entire system inherits the accuracy of its speech-to-text layer. This article maps the top CI tools by use case and explains why evaluating WER and diarization accuracy is the most reliable foundation for downstream insights.
Transcription quality: Key to CI performance
Speech and text analytics extract structured business signals from audio conversations: sentiment, topics, named entities, BANT qualifiers, and talk ratios. An AI notetaker records, transcribes, and summarizes meetings or calls automatically. Both categories depend entirely on the accuracy of the transcript they consume.
The sentiment analysis covered in this article is text-based inference from the transcript, where an NLP model analyzes the written words to determine emotional polarity. This differs fundamentally from acoustic emotion detection, which analyzes vocal characteristics in raw audio waveforms. We provide text-based sentiment inference, not acoustic emotion detection.
How transcription errors compound through analysis
A prospect says "I'm not interested in the enterprise plan." A transcription engine that drops the negation produces: "I'm interested in the enterprise plan." The sentiment model scores this as positive purchase intent. The sales team adds the lead to their hot pipeline, the CRM records a warm opportunity, and the coaching scorecard marks the rep's discovery as successful.
Every downstream system got it wrong because one word was missing at the first layer. The error doesn't fail loudly. It propagates silently through every system that consumed the transcript, and by the time anyone notices, the data corruption is already embedded in reports, forecasts, and coaching records. The same cascade applies to NER: a misheard company name or dollar figure produces a wrong CRM entry, a miscalibrated deal size, and a misleading pipeline report.
How WER is calculated
WER measures the minimum number of word-level edits (substitutions, deletions, insertions) required to transform a transcript into the correct reference, divided by the total words in the reference: WER = (S + D + I) / N. Transcription errors that enter the pipeline at the STT layer don't get corrected downstream, and the impact compounds through every CI technique that depends on what was said.
Lowering WER raises the ceiling on everything downstream
Solaria-3 is our most accurate model. Solaria-3 is our most accurate model for European business audio, covering English and core European languages (French, German, Spanish and Italian) across call center recordings that are noisy, compressed, and fast-paced. It ranks #1 against AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics on real customer recordings in English and core European languages, and achieves 6.4% WER on Earnings22 (the only model under 7%). Solaria-3 is async only, which reinforces the batch architecture this article recommends for CI workflows. For maximum language breadth, code-switching, or real-time streaming, Solaria-1 covers 100+ languages.
"Preferred vendor for speech-to-text speed & accuracy" - Verified user on G2
CI: From audio to actionable insights
Async (batch) transcription is the right architecture for most CI workflows. Processing the full recording before producing output gives the model complete conversational context, improving accuracy, speaker attribution, and multilingual consistency while still processing approximately one hour of audio in under 60 seconds. The following techniques all operate on async transcript output.
Assessing customer sentiment
Text-based sentiment analysis runs an NLP classifier over the transcript to assign a polarity score (positive, negative, neutral) to utterances or the call as a whole. The score is only reliable if the transcript accurately captured what was said, including negations, qualifiers, and speaker attribution. Our documentation covers implementation as part of the audio intelligence suite.
Prioritizing product roadmap topics
Topic modeling groups transcript segments by recurring themes, quantifying which feature requests, pain points, or competitor mentions appear most frequently across a call corpus. Accurate NER and consistent speaker labeling are prerequisites, and our Audio-to-LLM pipeline structures transcript output for exactly this kind of downstream analysis.
Extracting BANT from call transcripts
BANT extraction runs an LLM over the diarized transcript to identify Budget, Authority, Need, and Timeline signals. This is only reliable when the transcript correctly attributes each utterance to the right speaker, making low diarization error rate (DER) a hard prerequisite.
Pinpointing customer friction
Objection mining scans transcripts for competitor mentions, pricing objections, and feature gaps, surfacing patterns across hundreds of calls that individual review would miss.
Talk ratio: Measuring speaking time
Talk ratio measures the percentage of speaking time for each participant. Calculating this requires accurate speaker diarization, because the system must know which words belong to which speaker before it can measure their relative speaking time.
Mapping in-call emotional journey
Sentiment shift analysis tracks how emotional polarity changes from the opening to the close of a call. This technique requires accurate utterance-level timestamps and speaker labels to map sentiment to specific conversation moments.
Data-driven sales coaching with AI tools
Sales CI tools focus on rep performance: talk ratio, objection handling, and lead qualification accuracy. Each output depends on transcript accuracy and diarization precision. Automating CRM data entry from BANT extraction eliminates the manual post-call entry that sales reps routinely skip. Aircall, processing over 1M calls per week through us, uses this pattern to power CRM webhooks from a single API integration.
Automating support QA with AI analysis
Support CI tools apply transcript analysis to quality assurance: flagging escalations, categorizing tickets, and measuring resolution effectiveness. Automated QA can cover 100% of calls, but only if the underlying transcription is accurate enough to trust the outputs at scale. Gravite cut call quality review time by 93%, from roughly 15 minutes to 1 minute per call, while processing 50,000 hours of audio per year.
Transcription and sentiment QA tools
| Provider |
Core positioning |
Language coverage |
Diarization included |
Sentiment/NER pricing model |
Multilingual CI suitability |
Pricing structure |
Key QA consideration |
| Gladia |
End-to-end audio infrastructure for CI: record, transcribe, and enrich in a single API call |
Multilingual: 100+ languages |
Yes (async only, powered by pyannoteAI Precision-2) |
Included in base rate: diarization, sentiment, NER, translation, and summarization carry no add-on fees |
Solaria-3 leads on European business and contact-center audio (EN, FR, DE, ES, IT). Solaria-1 for broad language coverage and code-switching |
Starter $0.61/hr async. Growth as low as $0.20/hr async with upfront commitment |
No per-feature billing. QA cost scales linearly with audio volume |
| Deepgram |
Real-time and async STT, English-optimized |
English-optimized. Nova-3 Multilingual model available for select languages |
Yes (async only) |
Per-feature add-ons |
English-dominant pipelines. Nova-3 Multilingual extends coverage for select languages |
Per-minute, feature add-ons |
Strong for English-only QA pipelines. Limited coverage for multilingual contact centers |
| AssemblyAI |
Async STT with audio intelligence suite |
English-primary |
Yes (async only) |
Billed separately per feature |
English-only or English-dominant pipelines |
Base ~$0.15/hr async. Sentiment, NER, topics, and summarization each billed separately |
Per-feature billing compounds at scale when running QA across 100% of calls |
| Speechmatics |
Async STT, multilingual focus |
Broad European language coverage |
Yes (async only) |
Bundled in select tiers |
Strong European language coverage |
Per-hour, contact sales for full feature access |
Solid European language coverage. Narrower footprint for Southeast Asian or South Asian contact centers |
| ElevenLabs |
TTS-first platform with Scribe STT |
Multilingual |
Limited documentation |
Public pricing available for Scribe STT. Contact sales for audio intelligence features |
Limited CI-specific documentation |
Public per-hour pricing for Scribe STT. Audio intelligence features contact sales |
Primarily a TTS provider. Scribe STT has limited documentation for QA-specific workflows such as diarization and sentiment |
| Rev.ai |
AI transcription with optional human review |
English-primary. Foreign Language tier available at a separate rate |
Yes (async only) |
Separate pricing for human review |
English-primary workflows. Limited multilingual support via a separate Foreign Language tier |
Per-hour, tiered plans |
Human review improves accuracy on escalation-flagging workflows but adds turnaround time incompatible with same-day QA cycles |
Automated QA only delivers on its promise of 100% call coverage if the transcription layer is accurate enough to trust at that volume. Per-feature billing also compounds quickly at QA scale: running sentiment, NER, and topic detection across every call in a high-volume contact center adds a separate per-hour charge for each capability, and that cost structure multiplies directly with call volume. Our contact center as a service (CCaaS) infrastructure includes diarization, sentiment, NER, and summarization in the base rate, so QA cost scales linearly with audio volume rather than with the number of intelligence features enabled.
Uncover product gaps from customer calls
Sales and support calls contain product intelligence that most teams leave on the table. Feature requests, competitive objections, and pain point signals are distributed across hundreds of call transcripts. CI techniques applied at scale surface patterns that individual call reviews miss. Pairing topic modeling with sentiment scoring quantifies how strongly users feel about a specific missing feature, not only how often they mention it. Lost-deal analysis mines objection patterns for competitor mentions and the features that drove the switch, but this requires accurate speaker diarization to distinguish when the prospect mentions a competitor versus when the rep does.
How to evaluate transcription accuracy for your use case
A vendor claiming "100 languages supported" tells you nothing useful. What matters is WER on your specific audio conditions: phone calls, video conferences, bilingual teams, or noisy environments. Evaluating on your own data before committing to a provider is the strongest signal you can get: test your audio blind across six providers and pick the better transcript before you see the vendor name. For a folder of call recordings, the Gladia CLI runs a spot-check with --model solaria-3 and the diarization flag in a single command, before any integration code is written.
Testing WER on realistic audio across languages
Most STT benchmarks are measured on clean, studio-quality audio. Call center audio is different: compression artifacts, background noise, overlapping speakers, and accented speech all degrade accuracy in ways that clean-audio benchmarks don't predict. For call center audio, Solaria-3 ranks #1 against AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics on real customer recordings in English and core European languages, with open benchmark methodology. For multilingual coverage across 100+ languages or code-switching scenarios, Solaria-1 is the right model.
"Best multilingual real-time transcription on the market" - Yassine R. on G2
Diarization errors: Cost to product
DER measures diarization accuracy: DER = (Missed Speech + False Alarms + Speaker Errors) / Total Speech Time. High DER means the system misattributes speech to the wrong speaker, which corrupts every CI technique relying on speaker identity: talk ratio, BANT extraction, escalation detection, and rep-level coaching scores. Our diarization is powered by pyannoteAI's Precision-2 model, achieving on average 3x lower DER than alternative providers (Solaria-1, per the async benchmark) across eight providers. This is an async-only capability, not available in real-time workflows.
Ensuring production-ready accuracy
Published benchmarks are a starting point. Production references from comparable teams are more reliable. Claap, an all-in-one video workspace that helps organizations reduce meetings and messaging, reached 1-3% WER in production with one hour of video transcribed in under 60 seconds. Aircall cut transcription time by 95% (from 30 minutes to 1.5 minutes per call) and now processes over 1M calls per week through our API. Both outcomes are publicly documented and verifiable.
"Accurate Fast and Developer Friendly Transcription API for Multilingual Audio" - Faes W. on G2
Choosing and pricing your CI stack
Integration strategy: API vs. platform
Off-the-shelf CI platforms bundle transcription, analysis, and UI into a single product. That trades deployment speed for flexibility: the STT layer is fixed, so if it underperforms on your specific audio conditions, languages, or accent profiles, you cannot swap it out without replacing the entire platform.
Building on our API directly gives you full control: connect via REST for async workflows, route structured JSON to any LLM or analytics tool, and bring your own model or use our integrated options. Claap transcribed one hour of video in under 60 seconds, as documented in the meeting assistants case study.
Forecasting costs for your CI stack
Add-on pricing is where CI cost models break down. AssemblyAI's base transcription rate starts at $0.15/hr, but each audio intelligence feature adds a separate per-hour charge: Speaker Identification ($0.02/hr), Sentiment Analysis ($0.02/hr), Summarization ($0.03/hr), Entity Detection ($0.08/hr), and Topic Detection ($0.15/hr). Add PII redaction and content moderation and the effective rate reaches $0.68/hr or more. Deepgram's audio intelligence features are billed per token rather than per minute, adding projection complexity at scale.
We include diarization, sentiment analysis, translation, NER, summarization, and custom vocabulary in the base rate on Starter and Growth plans, with no add-on fees. At Starter pricing of $0.61/hr async, all features are included. Growth plans start as low as $0.20/hr with an upfront commitment.
At 10,000 hours per month, a five-feature AssemblyAI stack at $0.45/hr runs $4,500/month. Growth pricing at $0.20/hr runs $2,000/month. That's a $2,500/month delta, or $30,000/year, and the gap widens linearly with volume.
Start with €50 in free credit and have your async integration in production in less than a day. Test it on your actual multilingual audio, with your speaker count and your vocabulary, before committing at scale.
FAQs
Can I use one tool for both transcription and analysis?
Yes. We provide native recording, async transcription with speaker diarization, and a full audio intelligence suite (sentiment, NER, translation, summarization, chapterization) in a single API call. This collapses the typical 2-3 vendor stack into one integration point, eliminating the data degradation that occurs at each handoff between separate providers.
WER thresholds for sentiment precision
The threshold depends on use case. For decision-critical meeting and call recordings, where a missed negation or misattributed utterance corrupts CRM entries and coaching scores downstream, the downstream propagation mechanics described earlier in this article apply directly: a dropped negation flips a sentiment score. A misattributed utterance corrupts a BANT signal. The acceptable error rate is determined by the cost of acting on a wrong output, not by a universal threshold.
How do I handle code-switching in multilingual calls?
Traditional single-language detection identifies the primary language early in an audio recording and applies that model to the entire file, which fails when bilingual speakers switch languages mid-conversation. Our code-switching support detects mid-conversation language changes automatically across all supported languages, with stronger accuracy in async batch processing where full conversational context is available. This impacts operational costs when transcription breaks on bilingual calls.
Test Solaria-3 on your own business call audio, or Solaria-1 if your contact center handles multilingual or code-switching calls, using the async benchmark methodology to compare WER and DER against the specific alternatives you're evaluating. When the first layer of your CI stack produces accurate, diarized, multilingual transcripts, everything downstream works. When it doesn't, no amount of prompt engineering will fix what breaks at the source, and your engineering team spends sprint capacity patching around bad data instead of shipping product features.
What does our async transcription cost with all CI features included?
The Starter plan costs $0.61/hr async with diarization, sentiment, NER, translation, and summarization in the base rate and no add-on fees. Growth pricing starts as low as $0.20/hr async with an upfront volume commitment, and on Growth and Enterprise plans, customer audio is never used for model training with no opt-out required.
What WER does Solaria-3 achieve on business call audio?
Solaria-3 ranks #1 against AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics on real customer recordings in English and core European languages. On Earnings22 financial calls it achieves 6.4% WER, the only model under 7%.
How many languages do we support with code-switching?
We support 42 languages that no other API-level provider supports, in addition to broad coverage across all major language families. Code-switching detection works across all languages in both async and real-time modes, with higher accuracy in async batch workflows.
What is our real-time transcription latency for live agent assist?
Our real-time transcription delivers final transcript latency of approximately 300ms, making it viable for voice agent and live agent assist use cases. For post-call CI (sentiment, QA, and coaching analysis) async batch processing is the correct architecture: it costs less (as low as $0.20/hr vs. $0.25/hr real-time on the Growth plan), delivers better accuracy because the model has full conversational context, and processes approximately 60 seconds per hour of audio.
Key terms glossary
Word error rate (WER): The percentage of words in a transcript that differ from the ground truth, calculated as insertions, deletions, and substitutions divided by total reference words.
Diarization error rate (DER): A metric measuring how accurately a system attributes speech segments to the correct speaker, combining missed speech, false alarms, and speaker confusion errors. Lower DER is critical for multi-speaker meetings where follow-up action items need to be assigned to the right person.
Code-switching: The phenomenon where a speaker shifts between two or more languages within a single conversation or sentence. Most STT models trained primarily on English degrade significantly when speakers code-switch, producing silent accuracy failures that propagate into downstream summaries and follow-ups.
SOC 2 Type II: An audit certification covering security, availability, processing integrity, confidentiality, and privacy controls over a minimum six-month period, as defined by the AICPA. SOC 2 Type II is a voluntary attestation, not a legal mandate like HIPAA or GDPR. For B2B products processing customer audio, SOC 2 Type II is frequently a procurement and vendor contract requirement rather than a regulatory one.
