← All models

Azure Neural TTS

Audio Available Full comparison ↗

microsoft-azure/neural-tts · by Microsoft · neural-tts

Pricing — 1 offering(s)

Microsoft Azure

Standard voices

  • $4.00 / 1M characters Current 2019-09-01 → present
    Re-verified 2026-08-31 (SCO-521): rate unchanged at $4/1M, but non-neural Standard voices have moved to the…

    Re-verified 2026-08-31 (SCO-521): rate unchanged at $4/1M, but non-neural Standard voices have moved to the "Deprecated" section of the pricing page — "only available for existing customers to use." No end date announced; kept as an active price.

Neural voices

  • $16.00 / 1M characters Historical 2019-09-01 → 2026-08-30
  • $15.00 / 1M characters Current 2026-08-31 → present
    Re-verified 2026-08-31 (SCO-521): rate dropped from $16 to $15/1M. Pricing page now lists this as "Standard…

    Re-verified 2026-08-31 (SCO-521): rate dropped from $16 to $15/1M. Pricing page now lists this as "Standard Voice — Neural / Neural HD Flash (real-time and batch)". Exact change date not published; effective_from set to the verification date (was $16/1M when last verified 2026-06-17).

Custom Neural Voice (trained)

  • $24.00 / 1M characters Current 2021-01-01 → present
    Re-verified 2026-08-31 (SCO-521): unchanged at $24/1M — pricing page "Custom Voice — Professional Voice…

    Re-verified 2026-08-31 (SCO-521): unchanged at $24/1M — pricing page "Custom Voice — Professional Voice, Synthesis (real-time and batch)". Neural HD custom synthesis is a separate $48/1M rate.

HD voices (DragonHD / Neural HD)

  • $30.00 / 1M characters Historical 2025-03-01 → 2026-02-28
    DragonHD / Neural HD voices reached GA in March 2025 at $30/1M characters. Rate reconstructed from…

    DragonHD / Neural HD voices reached GA in March 2025 at $30/1M characters. Rate reconstructed from Microsoft's HD-voice update coverage — the live pricing page shows only the current rate. Superseded by the March 2026 reduction to $22/1M.

  • $22.00 / 1M characters Current 2026-03-01 → present
    Reduced from $30 to $22/1M characters in the March 2026 Azure AI Speech voice updates. Same Speech SDK / REST…

    Reduced from $30 to $22/1M characters in the March 2026 Azure AI Speech voice updates. Same Speech SDK / REST API as the Neural tier ("2.5" is a base-model version, not a separate product — SCO-465). The lighter DragonHD Flash tier is priced at the standard Neural rate ($15/1M) and is not broken out here.

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Azure Neural TTS fits into a cost-aware routing setup

See how →

Capability profile

How Azure Neural TTS rates across core capability dimensions, with the task-level evidence behind each rating.

voice naturalness Strong

High naturalness on neural voices; specific voices (e.g. Aria, Guy) consistently rated very natural in community comparisons. FILLED 2026-08-19 (SCO-462 sweep, was vendor/community-only): the Artificial Analysis Text to Speech Arena (artificialanalysis.ai/text-to-speech/leaderboard/provider-voice) ranks "Azure Neural" #70 of 98 voices tracked, Elo 1031 — solidly lower-mid-table, a more modest result than "consistently rated very natural" suggests. Note: the same arena also lists "Azure HD 2.5" at #23 (Elo 1132). Confirmed 2026-08-29 (SCO-465): "Azure HD 2.5" is Azure Speech's DragonHD / NeuralHD voice tier of this same TTS service (same Speech SDK/REST API — "2.5" is a base-model version), not a separate product. It is now broken out as this entry's `hd-voices` tier ($22/1M characters, GA March 2025, cut from $30 in March 2026 — SCO-506).

language support Strong

140+ languages and locales, 400+ neural voices — broadest language coverage among major cloud TTS providers.

voice variety Strong

400+ voices including speaking styles and role play capabilities. Custom Neural Voice allows brand voice training.

streaming latency Moderate

REST and streaming (WebSocket) APIs available. Latency is reasonable but not purpose-built for sub-50ms conversational AI.

cloning support Moderate

Custom Neural Voice (CNV) enables brand voice creation from recorded samples — a contracted enterprise feature, not instant cloning.

Benchmarks

Benchmark Score Config Source
MOS (naturalness, Azure en-US Aria — vendor-published) — Microsoft publishes internal MUSHRA and MOS scores favourably; no independent third-party MOS comparison published. —
Text to Speech Arena Elo (Artificial Analysis — "Azure HD 2.5" / DragonHD voice) 1132 #23 of 98 voices tracked — the independent Elo for this service's HD voice tier (now the `hd-voices` registry tier), confirmed same-product per SCO-465. Markedly above the "Azure Neural" voice (#70, Elo 1031). source ↗

Operator guidance

The natural choice for organisations running Azure-native infrastructure. At $16/1M chars (Neural) it is price-identical to Google Neural2 and Polly Neural, but Azure offers the widest language support (140+ languages) and speaking style controls. For voice cloning, ElevenLabs and PlayHT are simpler (no enterprise contract needed). For real-time conversational AI, Cartesia Sonic is faster.

Use cases

Limitations

Citations