---
title: "Gladia (Solaria)"
type: "AI Tool"
url: "https://aidemos.com/tools/gladia"
description: "Batch STT returned JSON word metadata and strong medical jargon recall, but crosstalk was incomplete and bilingual WER was skewed by duplicated channels."
category: "audio-speech"
website: "https://docs.gladia.io/"
published: "2026-08-13T09:18:22.552309+00:00"
updated: "2026-08-19T12:58:37.690031+00:00"
---

# Gladia (Solaria)

Batch STT with rich JSON metadata and strong jargon recall, but weak overlap handling and a channel-duplication caveat on bilingual audio.

## TL;DR Verdict

**Strong metadata and jargon recall, but not a clean win on hard overlap**

**Where it wins:**
- You need a batch STT API that returns transcripts plus word-level timestamps, confidence, and speaker labels.
- You care about technical jargon and proper-term recall on hard audio.
- You want mixed-language transcript coverage and can rerun stereo or code-switching cases with mono downmix or explicit channel control before trusting the WER.

**Main limitation:** You need reliable overlap handling on crosstalk-heavy audio.

**Pricing:** Starter (pay-as-you-go) Async $0.61/hr · Real-time $0.75/hr · Growth (annual commitment) Async as low as $0.20/hr · Real-time as low as $0.25/hr · Enterprise Custom

`Batch STT` · `Word timestamps` · `Speaker labels` · `Mixed-language`

**Website:** [Visit Gladia (Solaria)](https://docs.gladia.io/)

## Benchmark summary

Input 3's WER is not a trustworthy accuracy score because the transcript duplicated both channels.

- **37.35%** Overlapping Speech / Crosstalk — Four speaker labels were detected, but the transcript dropped 2,213 words and missed large stretches of dialogue.
- **4.07%** Medical Jargon — Best case: low WER and 100% recall on the scored jargon list.
- **invalid** Bilingual Code-Switching — The response duplicated both channels, so the WER is artefact-prone and must be rerun with mono downmix or explicit channel control before it can be cited.

> **Strong metadata and jargon recall, but not a clean win on hard overlap**
>
> Gladia is compelling as a batch STT API because it consistently returns word-level metadata and did especially well on the medical jargon sample. But the crosstalk transcript was incomplete, and the bilingual run's WER is not trustworthy because the response duplicated both audio channels; rerun with mono downmix or explicit channel control before relying on that score.

## Demo Recording

[Video: Gladia (Solaria) demo recording](https://cdn.futuresmart.ai/public/aidemos/1c791790364847e1b6def7578f202d80.mp4?v=1)
*Video — Screen-recording walkthrough of the benchmark workflow on a Mac desktop, moving through Finder, Terminal, and RStudio; it demonstrates the local evaluation process rather than the transcription result itself.*

## Feature-by-Feature Breakdown

### Batch Transcription

**Verdict:** Mixed

Accepts pre-recorded audio and returns transcripts in batch mode. The tested inputs included three long-form audio samples, showing the endpoint can process multiple recordings in one run.

**Input:**

> **Audio/Wav**

**Output:**

> **Application/Json**

**Input:**

> **Audio/Mpeg**

**Output:**

> **Application/Json**

**Input:**

> **Audio/Mpeg**

**Output:**

> **Application/Json**

**Bottom line:** Reliable as a batch endpoint, but transcript quality is uneven across hard audio and the bilingual score must be rerun before it can be cited.

### Structured Transcript Output

**Verdict:** Strong

Returns machine-readable transcript results with downstream-friendly metadata such as word-level timestamps, confidence values, speaker labels, and channel count. The tested runs on crosstalk, jargon, and bilingual audio showed the structured payload is consistently present and integration-ready.

**Input:**

> **Audio/Wav**

**Output:**

> **Application/Json**

**Input:**

> **Audio/Mpeg**

**Output:**

> **Application/Json**

**Input:**

> **Audio/Mpeg**

**Output:**

> **Application/Json**

**Bottom line:** Consistent, integration-friendly metadata export across every tested input.

### Mixed-Language Transcription

**Verdict:** Promising but caveated

Handles mixed English/Spanish audio without requiring a special code-switching flag. The bilingual sample produced Spanish-token coverage, providing evidence for mixed-language coverage.

**Input:**

> **Audio/Mpeg**

**Output:**

> **Application/Json**

**Bottom line:** Useful for mixed-language coverage, but this benchmark needs a rerun before the bilingual WER can be trusted.

### Speaker-Aware Transcription

**Verdict:** Mixed

Produces transcript output that preserves speaker labels and works on overlap audio while also continuing through bilingual runs. The tested overlap sample and bilingual runs showed speaker labeling and code-switching coverage in the transcript output.

**Input:** Overlapping Speech / Crosstalk

> **Audio** — Overlapping Speech / Crosstalk

**Output:** Run metrics

> **Image** — Run metrics

**Bottom line:** Useful for diarization-aware workflows and some mixed-language coverage, but speaker attribution correctness was not measured and the bilingual benchmark needs a rerun before its accuracy can be trusted.

## Vendor pricing

Starter matches the list-rate used in the benchmark; Growth lowers the per-hour price on commitment, and Enterprise is custom.

| Plan | Price | Notes |
| --- | --- | --- |
| Starter (pay-as-you-go) ★ (tested) | Async $0.61/hr · Real-time $0.75/hr | €50 free credits one-time; 25 async concurrent requests; 30 real-time concurrent requests; every feature and language included on paid plans. |
| Growth (annual commitment) | Async as low as $0.20/hr · Real-time as low as $0.25/hr | Requires sales contact; flexible concurrency; custom volume discounts; 99.9% uptime SLA; priority queue; model-training opt-out. |
| Enterprise | Custom | Requires sales contact; unlimited concurrency; zero data retention; SLAs; custom hosting; custom models; fine-tuning; debundled pricing. |

*Rates are from the vendor pricing page and differ between async and real-time usage.*

## Is It Right For You?

**Use it if**
- You need a batch STT API that returns transcripts plus word-level timestamps, confidence, and speaker labels.
- You care about technical jargon and proper-term recall on hard audio.
- You want mixed-language transcript coverage and can rerun stereo or code-switching cases with mono downmix or explicit channel control before trusting the WER.

**Skip it if**
- You need reliable overlap handling on crosstalk-heavy audio.
- You need a trustworthy bilingual WER without first rerunning with mono downmix or explicit channel control.
- You need measured streaming latency; this benchmark was batch-only.

## Classification

- **Category:** audio-speech
- **Subcategory:** audio-to-text
- **Type:** speech
- **Built for:** Other

## Frequently Asked Questions

**Q: Does Gladia return word timestamps and confidence scores?**

Yes. Across all three runs, the response payload exposed word-level timestamps, confidence, and speaker labels, with payload depth 3/3 and timed tokens present.

**Q: How accurate was Gladia on medical jargon?**

It did well. The medical jargon run scored WER 4.07% with 73 substitutions, 14 deletions, and 24 insertions, and it recalled all 9 scored jargon terms.

**Q: How did Gladia do on overlapping speech?**

Poorly. On the crosstalk sample it scored WER 37.35% and dropped 2,213 words, although it did detect 4 speaker labels that matched the 4 participants.

**Q: Is the bilingual WER trustworthy?**

No. The bilingual run carried two distinct channel values and duplicated the transcript across channels, so the 88.45% WER is an artefact rather than a valid accuracy score. It should be rerun with mono downmix or explicit channel control.

**Q: Did this benchmark measure streaming latency?**

No. This was a batch benchmark, so the measured latency is wall-clock batch processing time, not streaming lag.

**Q: What did Gladia cost in this benchmark?**

The benchmark used the vendor's list pricing. Starter async is $0.61 per audio-hour ($0.01017/min) and real-time is $0.75 per audio-hour; Growth and Enterprise are lower or custom on the vendor's pricing page.

**Q: Does Gladia handle mixed-language audio without extra configuration?**

The mixed-language sample showed Spanish-token coverage without a special code-switching toggle, with 56.2% Spanish token recall. However, the run is still not a trustworthy WER benchmark because the transcript duplicated both channels.

## Similar Tools

AI tools similar to Gladia (Solaria):

- [ElevenLabs](https://aidemos.com/tools/elevenlabs) — Natural-sounding voice cloning and narration, but with only approximate voice identity.

## Need a custom AI solution for this use case?

If you are looking to build a custom speech-to-text transcription, audio metadata extraction, or multilingual transcription system for your business or internal workflow, email us at [contact@futuresmart.ai](mailto:contact@futuresmart.ai).

### Found something inaccurate or missing?

We try to keep our AI research accurate and useful. If you found outdated information, an issue, or have a suggestion, email us at [collaborate@aidemos.com](mailto:collaborate@aidemos.com).