Gradium

Gradium is a low-latency text-to-speech vendor. It supports token-level streaming over a websocket, native telephony sample rates, and word-level timestamps, which makes it well suited to voice agents.

Get credentials

Sign in at gradium.ai.
Open API Keys and create a key.
Copy the value (it begins with gsk_).

Configure in jambonz

In the portal: Account → Speech → Add speech service → Gradium.

API Key

stringRequired

The key from Gradium’s console.

Model

string

default (recommended) or gradium-tts-beta. Defaults to default.

Gradium is text-to-speech only, so the Use for STT option is not offered.

Once saved, click Test to verify the credential. The voice list is fetched live from your account, so any custom or cloned voices you have created appear alongside the catalog voices.

Voices and languages

Gradium groups voices by bare language code rather than a locale — en, de, fr, es, pt. There is no en-US; use en.

1 {
2   "verb": "say",
3   "text": "Thanks for calling, how can I help you today?",
4   "synthesizer": {
5     "vendor": "gradium",
6     "language": "en",
7     "voice": "YTpq7expH9539ERJ"
8   }
9 }

The voice value is a Gradium voice id, which you can find in the portal’s voice dropdown or in Gradium’s voice library.

Voice settings

Vendor-specific tuning is passed in synthesizer.options. For Gradium this holds a json_config object plus an optional pronunciation_id.

1 {
2   "verb": "say",
3   "text": "Thanks for calling, how can I help you today?",
4   "synthesizer": {
5     "vendor": "gradium",
6     "language": "en",
7     "voice": "YTpq7expH9539ERJ",
8     "options": {
9       "json_config": {
10         "temp": 0.7,
11         "cfg_coef": 2.0,
12         "padding_bonus": -1.5
13       }
14     }
15   }
16 }

You can also set these in the portal under Extra Options on the speech credential, in which case they apply to every request that uses that credential. Settings supplied on the verb take precedence.

json_config

temp

float

Sampling temperature, 0.0 – 1.5. Defaults to 0.7. 0.0 is deterministic; higher values produce more varied delivery.

cfg_coef

float

Voice similarity. Defaults to 2.0. Higher values stay closer to the target voice; very high values can introduce artifacts.

padding_bonus

float

Speaking rate, -5.0 – 5.0. Defaults to 0.0. Negative values are faster, positive values are slower — note the direction is the opposite of most vendors’ speed controls.

rewrite_rules

string

Enables language-specific text rewriting applied before synthesis. Takes a language code, e.g. "en".

padding_bonus has a large effect on duration. On a short sentence, -4.0 produced 2.6 seconds of audio where the default produced 3.8 and +4.0 produced 9.3. Tune it before assuming a voice is simply slow.

Gradium validates temp and padding_bonus and rejects out-of-range values, but silently accepts out-of-range cfg_coef and unknown json_config fields. A typo will not raise an error — it will just have no effect.

pronunciation_id

string

Id of a pronunciation dictionary created in your Gradium account.

This goes at the top level of options, alongside json_config rather than inside it:

1 "options": {
2   "pronunciation_id": "your-dictionary-id",
3   "json_config": { "padding_bonus": -1.0 }
4 }

An invalid pronunciation_id fails the request with Pronunciation dictionary <id> not found.

Streaming

Gradium supports TTS streaming, so it can be driven token-by-token from an LLM using tts:tokens. Word-level timestamps are returned, so alignment events are available.

Two behaviours are worth knowing about when reasoning about latency:

Gradium has no cancel message, so a barge-in is handled by dropping and re-establishing the connection.
Completing an utterance ends the vendor’s session, so each turn uses a fresh connection.

Both are handled for you by jambonz; they cost a connection setup per turn, which is typically well under 300 ms.