A single mistranscribed witness name or garbled case citation in a deposition transcript can alter the course of a multi-million dollar lawsuit. Yet most legal tech platforms still evaluate transcription APIs on base price alone, ignoring three variables that determine whether the product holds up in production: accuracy on legal-domain audio, speaker attribution quality in multi-party proceedings, and whether the provider uses client audio to train AI models by default.
This playbook covers the technical and operational requirements for building AI transcription into legal and deposition workflows. We examine how to hit the word error rates required for defensible records, and how to handle speaker diarization without compromising attorney-client privilege.
Accuracy requirements for defensible legal records
The National Court Reporters Association (NCRA) requires 95% accuracy on each leg of the RPR certification exam, a pass mark that reflects the profession's baseline expectation for working reporters. In practice, production legal transcription workflows require human review to close the accuracy gap before any transcript is filed, with error rates varying significantly depending on audio quality, speaker count, and recording conditions.
Transcription precision for legal terms
Legal audio breaks most generic speech-to-text systems because it combines domain-specific vocabulary with real-world acoustic conditions: multiple speakers, overlapping speech, conference room reverb, and phone-quality recordings from remote witnesses. Common failure modes include Latin phrases like "voir dire" transcribed as "vore dear," case citation patterns misread, and legal acronyms like TRO (Temporary Restraining Order) corrupted into unrelated words.
Solaria-3 is built for exactly this audio profile: noisy, fast-paced, and multi-speaker. On Switchboard, conversational telephone speech, which matches the audio profile of remote witnesses dialing in from mobile phones, Solaria-3 scores 33.9% WER, the only model under 35% across comparisons against AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics. On Earnings22, which uses real financial and business speech, Solaria-3 achieves 6.4% WER, ranking #1 on that benchmark. Both figures translate directly into fewer corrections during human review of deposition transcripts.
International depositions introduce a separate set of requirements. When proceedings involve non-English witnesses, consecutive interpreters delivering the same testimony in a second language immediately after the witness, or parties switching languages during questioning, the transcription layer needs to track language changes without breaking speaker attribution. Solaria-1 is the right model for those workflows: it covers 100+ languages with native code-switching support, meaning it handles mid-conversation language shifts without requiring a separate detection pass or a second API call. The two models are complementary: Solaria-3 for English and core European business audio where accuracy on noisy, conversational recordings is the priority, Solaria-1 for proceedings that span languages or involve interpreters.
Cost of errors in depositions and court filings
Against those rates, automated transcription from service bureaus typically runs $0.10–$0.50 per minute, or $6–$30 per hour of audio, while professional human transcription runs $1.00–$3.00 per minute and specialized legal work higher still (per SpeakWrite's 2026 pricing guide). Our async rate runs $0.61 per hour on Starter and as low as $0.20 per hour on Growth, a difference of 10x to 150x against automated bureau rates, and substantially more against professional human transcription rates. That gap only materializes operationally, however, when the initial WER is low enough that editors spend their time on judgment calls rather than catching basic errors. An inaccurate first draft that requires line-by-line correction closes the cost gap quickly.
When transcription errors feed into downstream AI tools like summarization engines or case management indexers, the damage compounds further. A wrong party name in a transcript becomes a wrong name in every CRM entry, every AI-generated summary, and every search index built on top of it. These are errors that compound with each downstream system.
Target WER for legal audio workflows
For most legal transcription use cases, a human-in-the-loop review step before filing remains standard practice. Within the 3–8% WER range typical of production workflows, editors make corrections at a sustainable pace without extending turnaround times or adding headcount. At the lower end of that range, editorial review becomes a validation exercise rather than a correction exercise. At the upper end, editors are still catching roughly one error per 12 words, which is a meaningfully different workload.
The practical target depends on audio quality. A professionally recorded deposition with a dedicated microphone per witness differs substantially from a remote video call with participants dialing in from different devices. For depositions that include witnesses on mobile phones or in acoustically variable environments, Solaria-3's 33.9% WER on Switchboard is the directly applicable benchmark, not the clean-audio figures most providers lead with. It ranks #1 on that dataset and is the only model to score under 35% across AssemblyAI, ElevenLabs, Deepgram, Mistral, and Speechmatics.
How AI transcription handles speaker identification in depositions
Every Q&A exchange in a legal transcript must be correctly attributed to the examining attorney, the witness, or opposing counsel, because a diarization error that swaps speakers doesn't just create a confusing read, it creates a legally inaccurate document.
Diarization accuracy for multi-party proceedings
Our speaker diarization is powered by pyannoteAI's Precision-2 model, which is designed to handle overlapping speech in multi-party recordings. Diarization is strictly an async workflow, not available in real-time streams. The model analyzes the full recording before assigning speaker labels, which produces higher attribution accuracy than any real-time approach because it uses future context to resolve speaker ambiguity.
For legal tech platforms, this is the correct architecture: depositions are recorded first and processed second, and the accuracy trade-off strongly favors async processing.
Handling overlapping speech in depositions
In deposition settings, crosstalk occurs when attorneys interrupt witnesses, when multiple parties object simultaneously, or when a remote participant speaks while another is mid-sentence. The Precision-2 model handles overlapping speech explicitly rather than forcing a single-speaker assignment, which reduces the attribution errors that accumulate through a transcript when overlapping speech is treated as ambiguous.
For depositions with multi-channel audio, where each participant has a dedicated track, accuracy increases substantially because the model processes each channel as part of a unified workflow and assigns speaker labels at the word level across the full recording. Our audio-to-LLM pipeline processes multi-channel inputs and returns structured speaker labels across the full recording.
Standardizing speaker tags in AI legal transcripts
The async API returns structured JSON with word-level timestamps and speaker labels for every utterance. Below is an example of diarization output for a two-party deposition:
```json
{
"transcription": {
"utterances": [
{
"speaker": "speaker_0",
"start": 0.48,
"end": 4.21,
"confidence": 0.97,
"text": "Please state your full name for the record."
},
{
"speaker": "speaker_1",
"start": 4.85,
"end": 9.33,
"confidence": 0.94,
"text": "My name is Jonathan Reed, R-E-E-D."
},
{
"speaker": "speaker_0",
"start": 9.80,
"end": 14.10,
"confidence": 0.96,
"text": "Thank you. Were you present at the site on the morning of October 14th?"
}
]
}
}
```
The speaker field maps to unique speaker identities, the confidence score reflects the model's certainty on each utterance, and the start/end timestamps align with video timecodes for sync with recorded deposition footage. Legal tech teams can map speaker_0 and speaker_1 to named roles via their own application logic, then render the output in standard Q&A format for the final transcript. Teams building case management integrations can parse this JSON directly into their database schema, avoiding the fragile text-parsing logic required when working with unstructured transcript formats.
Confidentiality and attorney-client privilege in AI transcription
Under US federal common law and the Restatement (Third) of the Law Governing Lawyers § 79, voluntary disclosure to a third party who has no common interest in the privileged matter can constitute a waiver of attorney-client privilege. Several state bar ethics opinions, including guidance from the New York City Bar Association and the ABA's Model Rules commentary on confidentiality, have flagged the question of whether routing client communications through a provider that uses the data for its own commercial purposes is consistent with an attorney's duty of confidentiality, though courts have not yet resolved how that analysis applies to privilege waiver specifically. The law remains unsettled enough that legal tech platforms cannot rely on that ambiguity when their clients' communications are at stake.
Preventing model training on client audio
The distinction between our pricing tiers matters precisely here. On our Starter plan, customer data can be used for model training by default. On Growth and Enterprise plans, customer data is never used for model training, and no opt-out action is required. That default, not a contract clause buried in enterprise agreements, is what legal tech platforms need. The compliance risk isn't that someone forgot to opt out, it's that the team chose the wrong tier without knowing the policy.
"A reliable EU-based solution for transcription powered by AI" - Robin L. on G2
Retention and deletion controls
Legal tech platforms face conflicting retention pressures: courts may require transcripts for years in active matters, while data privacy regulations push toward minimizing how long third-party systems hold sensitive records. The correct architecture separates these concerns by retrieving transcripts immediately into your own compliant storage layer and triggering deletion of source audio from our infrastructure after verification.
Audit trails should log the API request ID, processing timestamp, deletion confirmation, and model version for each transcription to satisfy compliance and legal discovery requirements. Our PII (Personally Identifiable Information) redaction feature can replace sensitive identifiers like names and phone numbers with placeholder tokens in transcript output, but it requires explicit configuration per request and is not active by default.
SOC 2 Type II and data residency requirements
Our infrastructure holds SOC 2 Type II, ISO 27001, and GDPR certifications, with all customer data encrypted in transit and at rest. We operate dedicated infrastructure in EU and US regions, allowing legal tech platforms to route workloads appropriately to satisfy data residency requirements.
Measuring AI accuracy for legal transcription workflows
Benchmarking with real deposition data
Published benchmarks give you a starting point, but legal audio has characteristics that differ from standard evaluation sets: formal register, specialized vocabulary, and acoustic conditions ranging from professional studio setups to video call compression artifacts. Solaria-3 demonstrates strong performance on European business and conversational audio, evaluated across multiple datasets and hours of audio.
Use those figures as a floor, then benchmark against a representative sample of actual deposition audio with your own reference transcripts, that is the only evaluation that reflects your specific recording conditions and vocabulary. Our blind STT comparison tool works as a quick sanity check across providers before you run that evaluation. For an independently-run third-party ranking you can verify without running your own evaluation first, the Hugging Face Open ASR Leaderboard evaluates providers on a private conversational dataset outside any vendor's control. We rank #2 on that dataset, which is the kind of third-party signal a legal tech team can point to in an internal procurement review.
Measuring accuracy on case citations and proper nouns
Generic speech models have no prior knowledge of party names, case citations, or jurisdiction-specific terminology. For complex matters, load the case-specific terms, party names, and domain vocabulary relevant to that proceeding before submitting audio for transcription, and those are precisely the words where errors matter most. Custom vocabulary feature allows product teams to pre-load matter-specific terms, proper names, and citations before submitting audio for transcription, and it's included in the base rate on Starter and Growth plans.
Adapting speech recognition systems with in-domain vocabulary improves transcription accuracy over generic baselines, and the effect is most pronounced on the domain terms where courts focus most closely.
Higher initial accuracy directly reduces the time human editors spend on transcripts. A transcript at the upper end of the production range, around 8% WER, requires roughly one correction per 12 words. Push accuracy toward the lower end, around 3% WER, and that drops to one correction per 33 words, which changes the editorial task from active correction to light verification. Legal tech platforms that can deliver a 92–97% complete draft to their review layer achieve faster turnaround cycles and lower per-transcript labor cost, which is the operational model that makes AI-assisted legal transcription economically viable at scale.
Key considerations for legal transcription adoption
Legal admissibility of AI transcripts
In the United States, most state court rules and NCRA professional standards require that any transcript submitted as an official court record bear a certified court reporter's certification. AI-generated transcripts submitted without that certification and without human review would not satisfy those requirements in most jurisdictions, though the specific rules vary by state. Professional court reporting organizations provide guidance on electronic transcript authentication: Advisory Opinion 19 requires reporters to make a good faith effort to ensure the security of information contained in electronically transmitted transcripts. The practical implication for legal tech products is that AI tools function as efficiency multipliers for the certified reporter's workflow, not replacements for it. The correct integration model is AI-assisted: we produce a high-accuracy first draft, the certified reporter reviews and certifies, and the time saved per transcript translates directly into throughput capacity for the reporting agency. Legal tech platforms building for court-filing workflows should design their product architecture with this review step built in from day one.
Build vs. buy: self-hosted vs. managed API
The alternative to a managed API is self-hosting an open-source model. The table below compares the two approaches on the dimensions that matter most for legal tech product leaders:
| Evaluation criteria |
Self-hosted open source |
Our managed API infrastructure |
| Engineering resource allocation |
High, requires dedicated DevOps and ML engineers to maintain and scale |
Low, sub-24-hour integration using lightweight SDKs |
| Diarization quality |
Requires stitching together separate open-source libraries |
High, native pyannoteAI Precision-2 integration |
| Data privacy and sovereignty |
High if fully on-premise, but high infrastructure cost |
High on Growth and Enterprise plans (no model training, EU/US regions) |
| Total cost of ownership (TCO) |
Unpredictable due to GPU hosting and idle server costs |
Predictable per-hour billing: $0.61/hr async on Starter, as low as $0.20/hr on Growth. All features including diarization included in the base rate |
| File size limits |
Variable depending on hardware |
Up to 135 minutes and 1,000MB per file |
The leading open-source API option caps file uploads at 25MB and provides no native diarization, requiring teams to build and maintain a separate diarization pipeline. Teams self-hosting open-source models may encounter higher word error rates on production recordings without significant customization, which can push legal tech platforms back toward the manual review volumes they're trying to reduce.
Training custom vocabularies for legal terms
Custom vocabulary is included in the base rate on Starter and Growth plans. The accuracy improvement on loaded terms is consistent and measurable, and it's the most direct way to close the gap between AI accuracy on common vocabulary and AI accuracy on the case-specific language where courts focus most closely.
Start with €50 in free credits and have your integration in production in less than a day. Test Solaria-3 on your own multi-speaker legal audio to evaluate speaker diarization and accuracy on complex legal terminology before committing to a production architecture.
FAQs
Does your API use Gladia's legal audio to train its models?
On our Growth and Enterprise plans, your data is never used to train our models, ensuring complete confidentiality for sensitive legal proceedings. On the Starter plan, data may be used for training by default.
Is speaker diarization available for real-time deposition streaming?
High-precision speaker diarization powered by the pyannoteAI Precision-2 model is strictly available in our asynchronous (batch) workflow, not in real-time streams. For real-time workflows, speaker attribution should be handled in post-processing to guarantee accuracy.
What compliance certifications do Gladia's transcription servers hold?
Our infrastructure is certified under SOC 2 Type II, ISO 27001, and GDPR. We operate dedicated data residency regions in both EU and US.
What WER target should legal tech teams set for deposition transcripts?
The NCRA's RPR certification exam requires 95% accuracy to pass each dictation leg, a signal of the accuracy bar the profession considers baseline for working reporters, though it is an exam standard rather than a field requirement for delivered transcripts.
Does Gladia's API support non-English depositions or proceedings with interpreters?
Yes, through two models depending on the proceeding type. For depositions conducted in English or core European languages (FR, DE, ES, IT), including those with accented or non-native speakers, Solaria-3 delivers the highest accuracy on conversational and business-register audio. For proceedings that involve interpreters switching languages mid-session, non-English witnesses, or any workflow where the spoken language changes during the recording, Solaria-1 handles mid-conversation code-switching natively across 100+ languages. Both models return the same structured JSON output with word-level timestamps and speaker labels, so your application layer doesn't change between them.
How does multi-channel deposition audio improve diarization accuracy?
Our async API processes multi-channel audio and provides speaker labels in the final output. When each participant has a dedicated audio track, accuracy increases substantially because the model processes each channel as part of a unified workflow and assigns speaker labels at the word level in the final transcript. This significantly improves attribution accuracy over single-channel recordings with mixed speaker tracks, and it's particularly valuable for depositions recorded with a dedicated microphone per participant.
Key terms glossary
Word error rate (WER): The standard metric for transcription accuracy, calculated by dividing the sum of insertions, deletions, and substitutions by the total number of words in the reference transcript. Lower is better.
Diarization error rate (DER): The metric for speaker attribution accuracy, measuring the percentage of audio time where speaker labels are incorrectly assigned. Lower is better.
Speaker diarization: The process of partitioning an audio stream into segments by speaker identity, answering the question "who spoke when" across the full recording.
Code-switching: Mid-conversation alternation between multiple languages or dialects.
Attorney-client privilege: The legal protection that keeps communications between an attorney and their client confidential, which courts have held can be waived by sharing those communications with third parties that use the data for their own commercial purposes.
Data residency: The geographic constraint on where data is stored and processed, required by GDPR for EU-based data subjects. Our EU-west region satisfies this requirement without additional configuration.
