#

Key takeaways

• Nari Labs added word-level timestamps to its Qwen3-ASR endpoint.

• Clients can request the feature with a `word_timestamps` setting.

• The final transcript arrives first, then timing data follows in a separate event.

• Nari says the change is additive and does not remove existing fields.

• The company reports lower final-transcript latency in its own comparison.

What Nari announced

Nari Labs published an engineering note on September 23, 2026. The note announces word-level timestamps for its Qwen3-ASR speech-recognition endpoint. The feature is developer-facing and applies to the API layer, not to a new base speech model.

The company says a session can request the feature with a `word_timestamps` setting. The final transcript is delivered first. Word boundaries then follow in a separate event that uses the same item identifier. That lets clients connect timing data to the matching utterance.

How the timing data works

Each word includes a start time and an end time. Nari measures both in seconds from the beginning of the utterance. The company presents this as an additive change. It says the release does not remove or modify existing response fields.

The post also explains the internal split between recognition and alignment. The recognizer produces the transcript. A second model, Qwen3-ForcedAligner-0.6B, estimates word boundaries from the completed text and audio. In other words, the transcript and the timing pass are separate tasks.

Runtime design and latency

Nari says its runtime schedules speech recognition and alignment as independent tasks. It also batches compatible alignment jobs. The company says this protects the final-transcript path from being delayed by the aligner.

In Nari's comparison, final-transcript latency was 12 percent lower than a separate-process baseline while the aligner was running. This is a vendor-reported result. The source does not describe an independently run benchmark. It also does not publish the full test configuration.

What the comparison covers

Nari's own comparison of realtime transcription interfaces says several competing public schemas expose only segment-level timing or do not document per-word timing. The company says it checked that assessment against public API references and official SDK types on September 23.

The comparison excludes batch interfaces and undocumented features. That matters because the claim is narrow. It reflects the public developer-facing surface that Nari reviewed, not every possible implementation.

Operational considerations

The source suggests a few practical points for developers. First, the transcript arrives before the timing event. Clients therefore need to handle two related messages. Second, the timing data uses the same item identifier, so application logic must preserve that link.

Third, the feature is additive. Existing fields stay in place. That can reduce integration risk for clients already using the endpoint. Still, the source does not provide a full compatibility matrix, so teams would need to test their own workflows.

Morocco relevance

The source reports no Morocco-specific fact. A general lesson for readers is that additive API changes can be easier to adopt when they preserve existing fields and separate timing from transcript delivery.

Bottom line

Nari's update adds word-level timestamps to Qwen3-ASR through a separate alignment step. The design keeps transcript delivery moving while timing data follows afterward. The company also reports a latency improvement in its own comparison, but the source does not present an independent benchmark.

For developers, the main value is finer timing without replacing the existing response shape. For readers evaluating the release, the key limitation is clear: the source stays focused on the API feature and does not make broader claims about deployment scope, language support, or regional availability.