Speech provider study

This study compares local and hosted candidates for Slovak, Czech and English. The first declarations were checked on 21 October 2026, with the focused additions below checked on 23 October. It is not a language-quality bake or an admitted provider catalogue. Echo owns selection and local qualification owns the first experiment.

Local first

Treat the runtime and weights as separate selections. The runtime uses Apache 2.0, while each model has its own license; running locally does not make those weights unrestricted open source. Neither selected profile requires voice cloning.

Another local TTS candidate is OHF Piper, whose engine uses GPL-3.0. Its voice guide requires an ONNX model and JSON configuration per voice. Actual Slovak Lili and Czech Jirka medium cards each declare one speaker and 22,050 Hz. These are separate language assets, not one multilingual preset. This qualification did not install, select, time or listen to these Piper candidates.

Both cards label their target datasets CC0 but also declare fine-tuning from US English Lessac medium. Its card links research-use source-data terms. This study does not resolve the terms' effect on distributed weights or output; repository metadata alone cannot complete that provenance assessment. Engine, target dataset and model lineage remain distinct.

The first qualification pins audio.cpp release v0.9.2-drama3-base-small, commit 507495611b22dc4590e0e5609e05e8096384e33f. The CPU portable archive has SHA-256 d5e6110daae50800429680a277b70557d333d552c4a130a8f1d0e56181fa8697. This permits measuring speech without changing GPU drivers or allocating another GPU runtime. It is not a claim that CPU is the fastest backend.

Compatible standalone GGUF packages are pinned to audio-cpp/audio.cpp-gguf revision beb00ab19f10ca33e8a8983a43410527da3f5601:

The release model specifications define loaders and embedded assets. A similarly named GGUF for another engine is not interchangeable. The Parakeet runtime study supplies useful parity and precision evidence, rather than a measurement of this deployment.

Explicit CPU transcription alternative

The second candidate, checked on 22 October 2026, is whisper.cpp v1.9.5, pinned to d1be6fde11ac6e0407606b4e42fe72d34add8037, with multilingual ggml-small.bin. The MIT model is about 488 MB and has SHA-256 1be3a9b2063867b937e64e2ec7483364a79917e157fa98c5d94b5c1fffea987b. This is a separate transcription runtime; Supertonic synthesis can retain its existing audio.cpp binding. It is an explicit alternative, not an automatic fallback or a quality promotion.

The pinned server implementation accepts multipart file, language and response_format=json. Its startup --inference-path can select /v1/audio/transcriptions. The model loads at startup; a multipart model field does not select weights. The language table uses sk, cs and en. The engineering profile requires an explicit declared language rather than relying on the server's default English setting. A CPU build disables accelerator backends and starts with --no-gpu; short-command latency still needs local measurement.

Use the candidate to run a matched comparison before deciding whether to replace a recognizer. Give both recognizers the same immutable recordings, retain refusals, and test server text framing separately from word recognition. That separates a change in transport formatting from a change in what speech the model recognizes. State of Work owns measured delivery and the remaining human checks; this candidate alone cannot justify a global replacement.

Hosted alternatives

These are candidate dialects, not drop-in substitutes for audio.cpp. Hosted weights and service behavior are provider-managed. A language-identification feature does not by itself promise reliable code switching within one utterance. Vendor accuracy and latency figures do not qualify the target microphone, accent or game environment.

The ElevenLabs model matrix distinguishes file scribe_v2 from scribe_v2_realtime; measurements of one do not qualify the other. Flash v2.5 and Multilingual v2 list Slovak, Czech and English; a similarly named model is not interchangeable. Voice ID, model ID and language enforcement are separate selections. Google describes automatic recognition of the dominant language, while Azure continuous language identification explicitly does not detect word-by-word language changes within one sentence. Neither declaration establishes reliable Slovak/English game-term mixing. Keep single-language, between-utterance switching and within-sentence switching as different tests.

Select a paid service only after a matched recording comparison. Check current Azure region/tier pricing or ElevenLabs API pricing; no fixed cost or free-tier promise is assumed here. Credentials, destination, retention, voice consent and per-attempt cost bounds need their own admission. No purchase or hosted voice transmission is implied by this study. Google's separate processed-audio pricing also needs its exact method/SKU and channel accounting; no effective bill is inferred.

What to measure

Interleave Slovak, Czech and English recordings, then include one mixed-language case. Score intent, negation, item identity, count and coordinates separately from punctuation. Include stop, silence, clipped speech, player names and ordinary game noise. A TTS→STT round trip tests execution and transport, while independent human recordings and listener review test different boundaries. An ambiguous transcript must not become an inferred world action.

Measure cold and warm latency separately. Keeping the two small local models resident may avoid cold loading during play; thread count, idle unload and GPU placement still need host measurements. Optimize the whole path, including capture duration, GraphStep MapStep Map scheduling and game response, rather than reporting inference time as microphone-to-action latency.

When a command fails, distinguish three stages: the generated voice may pronounce the word poorly, the recognizer may misread it, or a correct transcript may fall outside the companion's declared fixed-word grammar. Preserve the original transcript and refusal so those causes remain investigable; do not map a similar-sounding word to a world action.

The current synthetic command qualification includes failed SK/CS/EN examples. Its TTS→STT round trips cannot identify one model as the cause or establish human accuracy. The next human comparison should therefore reuse exact recordings across recognizers and exact text across synthesizers. A listener scores pronunciation separately from parsing and local elapsed time.