Indian-language audio work frequently alternates scripts and English terminology. Sarvam’s Saaras model family exposes distinct modes that solve different tasks and should never be mixed up in a benchmark.
Why August 2026 matters
Sarvam’s August 18, 2026 changelog documents the introduction of saaras:v4 alongside saaras:v3 across REST, WebSocket and batch STT interfaces. It adds Global English in addition to Indian English and retains output modes for transcription, translation, verbatim rendering, transliteration and code-mixed text. The dated release note initially described v3 as the recommended default; later evolving documentation may display v4 as default. Report the model ID explicitly rather than claiming an unspecified default for the entire year.
Mode selection affects what counts as correct
A translation response cannot be scored against a same-language verbatim reference with ordinary word error rate. Transliteration may use a different script from the audio’s original language, and a code-mixed transcript can be better for a bilingual call center than an aggressively monolingual rewrite. Design separate golden datasets for each chosen mode. Preserve the input language, requested output form and any detected language uncertainty alongside the text.
Languages and audio transport
The provider lists a broad Indic-language family and language-code requirements using BCP-47-style identifiers. Coverage is not equal quality for every dialect or noisy phone line. The REST API reference describes special restrictions for raw PCM including the need for the correct codec declaration. Do not resample blindly and assume a wrong sample rate will be detected. API timeouts, buffering and diarization availability need endpoint-specific validation.
An editorial workflow
For a Hindi-English research interview, retain an encrypted source file, run transcription in a chosen mode, request a versioned language code and obtain a native editor’s check of names, quotes and number words. If a second transcript will become English subtitles, translate only after the source-language words are verified. Maintain both versions so a correction can propagate downstream. Treat inferred speaker identities as unverified unless supported by human review.
Which alternative to compare
Compare Saaras with Vosk or Whisper on your actual Indian language pairs rather than a general English leaderboard. A global hosted ASR product may offer broader enterprise integrations; Saaras may be attractive for output modes tailored to Indian scripts. Regional storage, enterprise retention, availability and actual per-minute prices are procurement questions; none should be inferred from a public marketing table.
Publication scope, rights and cost caveats
Research cutoff: 25 October 2026. The material above distinguishes provider documentation from independently measured behavior; no hands-on evaluation, legal clearance, voice-owner consent, commercial license purchase or model benchmark is claimed. Production teams must verify the specific API or installed version, account entitlement, input/output rights and destination market before release. Monetary plan amounts and metering tiers are not asserted here; the canonical pricing ledger belongs to AI 10. Missing release dates, exact rates and sample-rate capabilities are intentionally not inferred.